The AI Dividend and the Deadline — What Drives AI Tech Companies

Make AI Build AI. Get It in Everyone’s Hands. Offer Cheaply Today, But Not Tomorrow.

Abstract

AI companies build the models, chips and data centers that everyone else uses. This lens examines the industry from the inside: what these firms compete on, what limits them, and how they make money. The prize is large because everyone is realizing the value of AI and is buying more of it. This will drive market concentration. By 2028, the fastest few will own what the rest depend on.

• A1 — The Achievement: “Practical AGI”. The chapter sets out what AI can do today, using published measurements rather than claims. AI now matches or beats most people on most desk work. In 2025 both OpenAI and Google reached gold-medal standard at the International Mathematical Olympiad, a result only 67 of 630 human contestants matched. The leading models have also become hard to tell apart. On an independent ranking of overall model intelligence the top four sit within four points of each other, and one of those four is an open-weight Chinese model that anyone can download and run. Owning the best model is therefore a much smaller advantage than it was. Reliability remains the weak point. An agent works about eleven hours unsupervised if a 50% success rate is acceptable, but only about one hour at the 80% rate a business would deploy on.
• A2 — Assemble the Stack: Own What Is Scarce, Rent What Is Not. The chapter explains how AI firms decide which parts of the stack to build and which to buy in. They rent considerably more than their public positioning suggests. Anthropic pays about $1.25 billion a month for computing capacity inside a competitor’s data centers, while running more than a million of Google’s chips for its own service. The chapter also separates seven quite different businesses that all describe themselves as AI companies, from chipmakers to robot builders, and sets out the structural pressure each of them is under.
• A3 — Sustain the Capital Supercycle, Now Bound by Power. The chapter follows the money and the physical limits on it. The five largest US technology firms will spend about $765 billion during 2026, roughly three times their 2024 figure, and China is running a separate state programme worth about $295 billion over five years. Electricity has replaced chips as the binding constraint, and connecting a new data center to the grid can take five years. The opposition is now political as well as physical: US states introduced more than 300 data-center bills in the first half of 2026. Investors also pushed back for the first time, and Alphabet’s shares fell 7% on the day it raised its spending plans.
• A4 — Race the Frontier in a Two-Stack World. The chapter compares the American and Chinese AI industries, which stand closer together than most coverage suggests. Chinese models now carry about 61% of the traffic through the largest neutral model-routing service. Their prices are rising rather than undercutting, and one Chinese lab tripled the price of its newest model in a single generation. The chapter also covers a risk that is rarely priced in. In June 2026 a US export-control order required Anthropic, a leading American lab, to switch its newest models off worldwide, and they stayed off for 19 days.
• A5 — Capture the Customer, Then Convert It Into Revenue. The chapter explains how AI companies win users and then charge them. There are four ways to be paid, by subscription, by usage, by advertising and by enterprise contract, and the choice determines the risks. The sums involved are already large. The two leading labs run at roughly $74 billion and $41 billion of annual revenue, and advertising placed inside a chat assistant reached $100 million a year within two months of launch. What makes a customer stick to a service is the assistant’s stored comprehensive memory of that customer’s interactions, proving much more effective than website-specific browser cookies.
• A6 — Turn an Understanding of the Other Lenses Into AI-Company Advantage. The chapter looks at the same window from the four other positions in this paper: the businesses that buy AI, the people who use it and build it, the governments that fund and restrict it, and the world system that decides where the profit finally settles. It draws out the moves an AI company can only see from outside its own industry, including the distribution that European regulators are increasingly forcing open.

Table of Contents

• A1. The Achievement: “Practical AGI”
  • A1-1: Read the scoreboard
  • A1-2: Capability comes from the agent loop, not any single model
• A2. Assemble the Stack: Own What Is Scarce, Rent What Is Not
  • A2-1: Fifteen companies hold the AI stack, and NVIDIA
  • A2-2: “AI company” names seven different businesses, and they do not share economics
  • A2-3: What actually drives each of the seven types
• A3. Sustain the Capital Supercycle, Now Bound by Power
  • A3-1: The constraint moved from chips to electricity, and the capital is borrowed against it
• A4. Race the Frontier in a Two-Stack World
  • A4-1: AI improving AI is the engine under the frontier race, and its doubling times are measured in months
  • A4-3: Two cost curves move in opposite directions, and both are true at once
  • A4-4: Two ways to build an AI industry
• A5. Capture the Customer, Then Convert It Into Revenue
  • A5-1: Four ways an AI company gets paid, and each one buys a different kind of customer
• A6. Turn an Understanding of the Other Lenses Into AI-Company Advantage

Preface

This is a private synthesis of what matters in AI in mid 2026.

• Initiative: this is a personal, private project and freely shared.
• Intent: this distills what matters in AI in mid 2026 from a 360-degree perspective. Emphasis is to be quantitative even when uncertain. The aim is on how to reap the AI dividend while containing the associated risks.
• Source: initially built on systematic evaluation of approximately 19,000 AI-oriented news articles published in the past four months, supplemented by focused additional research where the news flow needed depth, checking, or quantitative grounding. AI itself was used, of course, to help assemble, structure, and refine the synthesis.
• Style: clear statements in a clear structure rather than continuous prose.
• Caveat: compiled in a few weekends. Please accept this as an imperfect, focused big-picture synthesis now, rather than a detail-honed article six months from now. The 2026–2028 window does not wait.
• Disclaimer: shared for information only. No advice. Without warranty.
• Feedback: is highly welcome, e.g. on my Substack channel

↑ Back to Table of Contents

Reader Overview

This document reads the 2026–2028 window through the AI-company lens (A), the layer that builds the capability everyone else deploys. Five stakeholder lenses map the whole window: AI companies (A), businesses (B), individuals (I), governments (G), each racing their own rivals, and the world system (W). This view carries Lens A. The other four stakeholder lenses are already reinterpreted from the AI-company perspective in the final chapter (A6). One warning for the general reader: “AI company” is not one business. Seven distinct types sit on this stack, and breakout A2-2 separates them before the rest of the argument depends on the difference.

↑ Back to Table of Contents

A1. The Achievement: “Practical AGI”

Summary: “Practical AGI” has arrived. AI now matches or beats most people on most cognitive work: broad knowledge tests at ~90%, gold at the 2025 International Math Olympiad (35/42, beaten by all but 67 of 630 humans), coding from ~5% to ~90%+ on the standard coding benchmark. Academics still reserve the term for ~2047. The version that reorders work is already here. By July 2026 the frontier had converged into a four-point band, with an open Chinese model inside it, so owning the best model buys little. AI capability now lives in the agent, different models in a plan-act-observe loop, not in any single model. Reliability is the gap that remains: agents sustain ~11 hours unsupervised at a 50% success rate, but roughly an hour at the 80% rate anyone would ship on. The competition runs on all fronts: peak skill, agents, reliability, and cost.

Evidence:

• The claim rests on measured results, not impressions. Across every modality with a public measure — reasoning, text, images, music, video, voice, code and paid economic work — frontier systems now match or beat most people. Breakout A1-1 puts the scoreboard in one place.
• The frontier is a dead heat, and one of the leaders is downloadable. On the Artificial Analysis Intelligence Index, Claude Fable 5 scores 60, GPT-5.6 Sol 59, Moonshot’s open-weight Kimi K3 57 and Claude Opus 4.8 56 (The Decoder). K3 carries 2.8T parameters, a 1M-token context and open weights, and leads a frontend-coding arena ahead of both closed rivals.
• The pace matters more than the level. On the Remote Labor Index, which scores whether agents finish real paid freelance projects at a quality a client accepts, the frontier more than quadrupled in under eight months, from 4.17% to 16.1% (CAIS). That is still only one project in six.
• Applied domains confirm it outside the benchmarks. Legal advice, protein-structure prediction (AlphaFold, 2024 Nobel) and autonomous driving (Waymo, Baidu Apollo Go) run on the same capability, taken up in Lens B.

• Breakout A1-1: read the scoreboard — AI already matches or beats most people in every modality that is measured.

  One claim carries this chapter, so here is the evidence for it in one place: one row per capability, each a measured result with a source. No single row settles the question. Read down the column and the pattern does.

  Capability	Achieved proficiency
  Reasoning	Expert level. Knowledge tests sit at MMLU ~90%, and on GPQA Diamond, where PhD experts score ~65%, models cleared the bar in 2024 (Stanford HAI). In 2025 both OpenAI and Google reached gold-medal standard at the International Mathematical Olympiad (35/42), matched by only 67 of 630 human contestants (DeepMind). Chinese labs sit at parity
  Text	Out-argues people. In a randomized trial GPT-4 beat human debaters ~64% of the time, and ~81% more often when given a few personal facts about the other side (Nature Human Behaviour)
  Images	Indistinguishable from real. AI-synthesized faces can no longer be told from photographs, and are rated more trustworthy (PNAS)
  Music	Most of the new supply at one service. By April 2026 ~44% of tracks uploaded to Deezer daily (~75,000) were fully AI-generated, up from ~10% a year earlier (TechCrunch)
  Video	Photoreal clips with synced dialogue and sound: OpenAI’s Sora 2 (OpenAI). China leads on scale, Kuaishou’s Kling at 60M+ creators and revenue up ~300% YoY (SCMP)
  Voice	Conversational tempo. ~320 ms speech-to-speech response against the ~200 ms gap people leave between turns (OpenAI)
  Code	Near-saturated. On SWE-bench Verified, scores climbed from ~5% (2023) to ~90%+ (2026) (Epoch AI). China’s open-weight GLM-5.2 now tops the harder SWE-Bench Pro, ahead of GPT-5.5 (VentureBeat)
  Economic work	The weak row, and the one that pays. The Remote Labor Index scores whether agents finish real paid freelance projects at a quality a client accepts. Anthropic’s Fable 5 completed 16.1% in mid-2026, double Opus 4.8 (8.3%) and ahead of GPT-5.5 (6.3%) (CAIS), having more than quadrupled in under eight months from Opus 4.6’s 4.17%

  • The pattern across the rows carries the argument. Different labs, different modalities, different measures, all moving the same way inside three years. A weak result in any one row does not undo the sweep.
  • The weakest row is the one that matters commercially. Reasoning, text and code are saturated. Finished economic work sits at 16.1%. The distance between those two numbers is the reliability problem, and it is why a person still signs off (breakout A1-2).

• Breakout A1-2: capability comes from the agent loop, not any single model — and it is real and compounding.

  Capability above no longer comes from one model answering a prompt. An agent wraps models in a loop (plan, act, observe, repeat), routing each step to whatever fits: a cheap model for easy steps, a heavy reasoning model for hard ones, a tool for the rest. The combination is more capable than any single model, and it is what the 2025–26 agent wave runs on. Six elements make the loop, and each one adds capability a single model call cannot reach.

  Loop element	What it does	What it adds over one model call	Measured today
  Plan	Break a goal into ordered steps, and revise the plan as results come back	A single call answers. A plan lets the system attempt work that has no single answer	Agents sustain ~11 h of unsupervised work, the 50%-task horizon doubling every ~89 days (METR)
  Act (tools)	Call other software, run code, query databases, operate applications	Turns a text predictor into something that changes state in the world	Open tool-access standards (MCP and rivals) are now the default integration path (A5)
  Observe and verify	Read the result, test it, and retry when it fails	Errors get caught inside the loop instead of shipping	Verification is what separates the ~5% of enterprise deployments that scale (B3)
  Remember	Carry state, findings, and user context across steps and sessions	Work survives beyond one context window, and the system improves per user	Agent memory is now the main defensibility in consumer AI (A5)
  Route	Send each step to the cheapest model that can do it	Cost falls by an order of magnitude without losing peak capability on hard steps	Realized ~$0.99/M tokens against a $5/$25 sticker, on >90% cache hits (A5)
  Human gate	Hold the costly and irreversible steps for a person to approve	Keeps a confident error from becoming an expensive one	The best agent still finishes only 16.1% of real paid freelance projects at client-acceptable quality (CAIS)

  • The loop is where the last two years of capability actually came from. Base models improved, but the step change in what AI can finish came from wrapping them: coding agents author ~35% of merged pull requests at one major tool (Cursor), and Anthropic reports AI is involved in >80% of the code merged inside the company (Anthropic).
  • Reliability, not skill, is the binding element. Five of the six elements raise capability. The sixth exists because the other five are not trustworthy enough to run unattended on anything expensive. That is the whole reason “agents will just do the work” runs ahead of what ships.
  • The loop is buildable by anyone, which is why it is not a moat. Every element above is public engineering, not a secret weight. The advantage goes to whoever runs the loop best on data nobody else has, which is why A5 argues the fight is over memory and orchestration rather than the model.

  The hype runs ahead of reliability, but the capability under it is real and compounding.

What This Means:

• The label is debated. The practical reality is not. “AGI” in the academic sense (smarter than every person at every task) is still distant. A survey of 2,778 researchers puts 50% odds only around 2047 (Grace et al.), and on Humanity’s Last Exam, the benchmark built to be expert-hard, the best models still score only ~50% (HLE). Microsoft called GPT-4 an early AGI (arXiv) and DeepMind ranks today’s systems “emerging AGI” (arXiv). The version that matters in real life has arrived: AI is smarter than most people on most tasks. It is that “practical AGI”, not the academic finish line, that reorders work, business, and power.
• AI capability is now a property of the agent, not of any one model. The frontier is systems that stitch many models and tools together. Owning “the best model” matters less than building the best agent around it, and July 2026 made that concrete: four points separate the top four, so the differentiator moved to the system built on top. Kimi K3 leads a coding arena while ranking fourth on general intelligence, which is what a commodity band looks like from the inside.
• The US and China stacks differ in kind, not just speed. The US sells closed, premium models through APIs, services split across many firms. China folds open-weight models priced 5–30× cheaper (Artificial Analysis) into super-apps that already own messaging, payments, and shopping. China reaches deployed scale faster while the US holds the capability lead (A4).
• Progress runs on three fronts at once: peak skill, reliability, and cost. Peak skill keeps rising (e.g. software coding), cost keeps falling, and reliability is improving too. Reliability still trails the other two, and METR’s two horizons size the gap: the agents that sustain ~11 hours of work at a 50% success rate hold only about an hour at the 80% rate a business would deploy on (METR). Almost every deployment decision sits in that gap. Models fabricate fluently with no calibrated sense of their own uncertainty. Security and guardrails vary widely across labs (G3, G4). AI also fails differently from people: its errors look like competence, fluent and confident, so they slip past checks built for human mistakes.

Recommendations:

• Compete on frontier models, agents, and cost, the three fronts that still move. The edge is in orchestration, tools, evaluation, and the proprietary data wrapped around models, plus the frontier capability and unit cost that decide who can afford to serve it. With four points separating the top four models, betting the company on a raw capability lead is betting on the narrowest margin in the industry.
• Spend on the reliability layer first. Evaluation, monitoring, provenance, guardrails, and security decide what you can actually ship. Capital still underprices it relative to raw capability.
• Match autonomy to stakes. Use full autonomy on low-risk, reversible work. Keep a human on the approvals wherever a confident error is expensive.
• Everything else is the rest of this lens: integrate across the stack (A2), sustain the capital cycle (A3), race the two-stack frontier (A4), and capture the customer and agent layer (A5).

↑ Back to Table of Contents

A2. Assemble the Stack: Own What Is Scarce, Rent What Is Not

Summary: The stack is assembled, not simply owned. Integration still pays where a layer is scarce. Hyperscalers ship their own silicon, from Google Ironwood to AWS Trainium3, and China’s super-apps fold chips, cloud, models, and payments into one closed ecosystem. Everywhere else 2026 turned to renting, including from direct rivals: Anthropic pays xAI about $1.25B a month for capacity and offered Meta up to $10B, while Meta weighs selling compute at all. The phrase “AI company” also hides seven different businesses, each with its own economics and its own structural squeeze (breakouts A2-2, A2-3). No stack closes fully, and the upstream limit has moved: advanced packaging and high-bandwidth memory now bind past 2030, not 2027. Custom silicon crossed from experiment to scale in 2026, with over a million Google chips serving one lab’s inference. Own the one or two layers where scarcity justifies it. Rent the rest on terms you can exit.

Evidence:

• Custom silicon crossed from experiment to scale in 2026, and the reason is unit cost. Anthropic now runs more than one million Google Ironwood chips for Claude inference, the first custom accelerator at seven-figure volume with a single customer, and has committed to up to 5 GW of AWS Trainium capacity; Amazon had shipped ~1.4M Trainium2 chips by early 2026. Meta’s MTIA, co-designed with Broadcom and built by TSMC, entered production for a September 2026 start, and Apple committed ~$30B with Broadcom to co-design chips onshore (Tom’s Hardware). Reported migrations off merchant AI chips onto custom accelerators cut inference bills by roughly two thirds, which is what pays for a silicon programme. Custom silicon is now a baseline requirement for the hyperscalers. The pure-play labs mostly rent. The big clouds ship their own accelerators at scale (Google Ironwood Google’s own AI chips and AWS Trainium3 both now generally available, Meta MTIA 300, Microsoft’s next-gen Maia 200 slipped to 2026, Apple’s Baltra reported for ~2027), with Anthropic alone committed to up to ~1M Google chips (Anthropic). The frontier labs are the opposite. OpenAI’s first chip (Jalapeño, with Broadcom) is inference-only, unveiled June 2026 and rolling out from late 2026 (TechCrunch). Anthropic owns no silicon, xAI runs NVIDIA, Mistral is only exploring, and Tesla cancelled Dojo in 2025 (TechCrunch). Owning a chip is a hyperscaler edge the labs lack. China’s answer is Huawei Ascend at frontier scale (A4).
• Foundation labs integrate downward into compute and energy: OpenAI Stargate, Anthropic’s multi-year cloud contracts, Mistral’s NVIDIA-GB300 facility in Paris.
• Model development has separated from infrastructure ownership, and labs now rent from rivals. Anthropic runs on xAI’s Colossus 1 at ~$1.25B a month through May 2029, terminable on 90 days’ notice, holds a 20-year lease on ~401 MW from TeraWulf worth ~$19B of contracted revenue, committed ~$50B to Fluidstack for custom sites in Texas and New York, and as of 17 July 2026 had offered Meta up to $10B of capacity over two years (Forbes). Meta is weighing a cloud business to sell its surplus. Advantage now goes to whoever assembles power, chips, interconnection, and permits fastest, not to whoever owns the most.
• The Chinese alternative stack runs at frontier scale. Huawei Ascend + the CANN software stack power Z.ai’s GLM-5, a frontier-tier model trained entirely on Huawei (non-US) hardware (SCMP). Chinese vendors reached ~41% of China’s AI-accelerator market in 2025 (IDC).
• Chinese super-apps integrate every layer in one closed ecosystem: Alibaba (chips + cloud + Quark/Qwen + Taobao/Alipay), ByteDance (Doubao + Douyin), Tencent (Yuanbao + WeChat 1.4B) (SCMP).
• The upstream limit stretched from 2027 to past 2030, and it moved from packaging to memory. TSMC names advanced packaging rather than wafers as the true bottleneck, and SK Hynix’s chief executive now says 2027 will be the worst supply shortage in memory’s history, with demand exceeding capacity beyond 2030 (Astute Group). One NVIDIA Rubin accelerator needs eight next-generation high-bandwidth memory stacks, so a handful of memory suppliers, not the chip designers, now set how many accelerators reach the market. NVIDIA answered in July 2026 with a >$500B strategic arrangement with SK Group, including long-term next-generation high-bandwidth memory supply for the Vera Rubin generation. Even integrated players hit upstream limits: advanced chip packaging at TSMC is booked into 2027 and 2026 high-bandwidth memory output is sold out (CNBC). China’s “independent” stack is capped the same way. Domestic high-bandwidth memory holds Ascend output far below demand (SemiAnalysis), the binding constraint on its self-sufficiency drive.

• Breakout A2-1: fifteen companies hold the AI stack, and NVIDIA — selling to all — is worth the most (approximate snapshot, mid-2026).

  Figures are rounded and approximate. Market caps move daily, private valuations are the latest round, and profit and user counts mix fiscal years and definitions. Read this as a map of who holds what, not as audited accounts.

  Company	Value (mid-2026)	Users	Latest annual profit	AI spend (2025–26)	Major non-AI business
  NVIDIA	~$4.8T market cap	— (infrastructure)	~$75–90B net	R&D + ecosystem stakes	data-center networking, gaming GPUs
  Apple	~$3.3T	~2.3B active devices	~$100B net	modest; inference chip ~2027	iPhone, services, devices
  Microsoft	~$3.7T	Copilot/365 in billions of seats	~$100B net	~$190B FY26 capex	Windows, Office, Azure, Xbox
  Alphabet	~$3T	Search billions; Gemini ~650M	~$110B net	~$185B 2026 capex	Search ads, YouTube, Android
  Amazon	~$2.5T	Prime ~200M+	~$60B net	~$200B 2026 capex	e-commerce, AWS, advertising
  Meta	~$1.8T	~4B family-of-apps	~$65B net	~$135B 2026 capex	social apps, advertising
  Anthropic	~$965B (private; filed 1 Jun 2026)	Claude, tens of millions	loss (~$47B run-rate rev)	>$100B AWS + ~1M Google TPUs	none (pure AI)
  OpenAI	~$852B (private)	ChatGPT ~900M weekly	loss (~$20–25B ARR)	~$50B 2026; Stargate $500B multi-yr	none (pure AI)
  Oracle	~$0.7–1T	— (enterprise)	~$12B net	~$35B 2026 capex; $638B backlog	databases, enterprise apps
  xAI	~$200B+ (in SpaceX; bought Cursor $60B)	Grok via X, Cursor	loss	$20B raised; Colossus ~1M GPUs	(SpaceX) launch/Starlink; (X) social
  Tencent	~$550B	WeChat 1.4B; Yuanbao	~$30B net	large multi-year capex	games, WeChat, fintech
  Alibaba	~$320B	Quark 151M; Alipay ~1B	~$15B net	RMB 380B (~$54B) over 3 yr	e-commerce, cloud, payments
  ByteDance	~$330B (private)	Douyin/TikTok ~1.5B; Doubao 150M	~$30B+ net	large; Doubao + own silicon	short-video, advertising
  Huawei	private (~$100B+ revenue)	— (devices/telecom)	profitable	Ascend chips + CANN stack	telecom gear, smartphones, cloud
  DeepSeek	private (High-Flyer-backed)	DeepSeek app (top-ranked)	n/a	~$1.6B GPU fleet (est.)	quant hedge fund (High-Flyer)

  Two patterns stand out. The hyperscalers and Chinese super-app giants fund the buildout from huge non-AI profits, while the pure-play labs (OpenAI, Anthropic) carry the highest valuations on losses and outside compute. Anthropic’s run-rate revenue shows the pace: $9B (Dec 2025) → $14B (Feb) → $30B (Apr) → $47B (May 2026) (Simon Willison, from Anthropic disclosures). NVIDIA, selling to all of them, is worth more than any. Sources: market caps companiesmarketcap, Anthropic ~$965B Series H (Anthropic), OpenAI ~$852B and ChatGPT ~900M weekly (A3 and OpenAI), hyperscaler capital spending per A3, Chinese-firm figures per A2/A4.

• Breakout A2-2: “AI company” names seven different businesses, and they do not share economics.

  The seven types are cut by core value creation, so each row is one business line, not one company. A firm with several lines appears in several rows: Amazon sits in compute landlord for AWS and in distribution platform for retail advertising, Google in chipmaker, compute landlord, frontier lab, and distribution platform at once. That is what makes the split exhaustive without counting one company twice. The revenue column does not sum to a market total, because each layer’s revenue is the next layer’s cost: a lab’s token revenue pays a compute landlord, whose rent pays a chipmaker. Revenue is the AI-attributable part only, for 2026, and is modeled from the cited market and company figures.

  #	Type	Core value creation	AI revenue 2026 (modeled)	Key players (global)
  1	Chip and hardware makers	Design and sell the silicon, memory, and networking that training and inference physically run on	~$281B of “intelligent” data-center silicon inside $477B data-center semiconductors (IDC)	NVIDIA (~87% of merchant AI-chip revenue), Broadcom (~$56B custom silicon), AMD, TSMC, SK Hynix, Micron, Samsung, Huawei
  2	Compute landlords	Rent capacity by the hour or on 20-year contract; carry the capital and power risk	~$100–150B, hyperscaler AI cloud plus neoclouds (CoreWeave ~$8B run-rate, $99.4B backlog)	AWS, Microsoft Azure, Google Cloud, Oracle, CoreWeave, Nebius, Fluidstack, Alibaba Cloud
  3	Frontier model labs	Train and serve general-purpose models; sell tokens, subscriptions, and licences	~$90–110B run-rate (Anthropic ~$47B, OpenAI ~$25B+, plus Google, xAI, Chinese labs)	OpenAI, Anthropic, Google DeepMind, xAI, Meta AI, Mistral, DeepSeek, Alibaba Qwen, Moonshot, Z.ai
  4	Application and agent firms	Turn a model into a finished job inside one workflow; sell the outcome, not the model	~$25–30B, led by coding and legal (Cursor ~$4B annualized, Harvey ~$300M, Sierra ~$150M)	Cursor (SpaceX), Harvey, Sierra, Glean, Abridge, Cognition, Clio
  5	Distribution platforms	Own the user relationship, inject AI into it, and monetize the attention through ads and subscriptions	~$35B+ AI-attributable advertising, rising toward ~$142B by 2030	Google Search, Meta, Microsoft 365, Apple, Tencent, ByteDance, Alibaba
  6	Data, tooling and evaluation	Supply the training data, pipelines, evaluation, and observability the other six depend on	~$26B data-science and machine-learning platforms (Gartner)	Databricks, Snowflake, Scale AI, Surge, Hugging Face, LangChain, Weights & Biases
  7	Embodied AI and robotics	Put models into machines that act physically; sell the machine and the autonomy on it	~$5–10B today, ~$38B by 2035 (Goldman Sachs)	Tesla Optimus, Figure, Unitree, Agility, UBTech, Fourier, LimX

  • The types differ in what they actually sell, so they differ in what can destroy them. A chipmaker is destroyed by a process node it cannot get. A compute landlord is destroyed by a funding freeze against 20-year debt. A lab is destroyed by a rival’s open weights. An application firm is destroyed by the lab it builds on shipping the same feature. Reading them as one industry hides all four risks.
  • Margin sits at the ends, not the middle. The chipmaker at the top prices against a bottleneck and the application firm at the bottom prices against a delivered outcome, so both hold margin. The compute landlords in between carry the heaviest capital and the thinnest cover, which is why a correction prunes type 2 first (A3).
  • Type 5 is the quiet giant. Distribution platforms had the users before AI existed, so they monetize AI without having to win a model race. That is the position the labs are buying their way toward, and the reason advertising arrived in chat assistants in 2026 (A5).

• Breakout A2-3: what actually drives each of the seven types — the pressure each one is under, and what it does about it.

  Breakout A2-2 says what each type sells. This one says what keeps each of them awake. The pressure column is the structural squeeze the type cannot escape by working harder. The action column is what firms of that type are visibly doing about it in 2026, which is why several of them are moving into each other’s territory.

  #	Type	Strategic challenge	Resulting action in 2026
  1	Chip and hardware makers	The contest moves from peak training performance to cost per inference, where specialist silicon can undercut a general-purpose part. Customers are also becoming competitors as hyperscalers design their own	Defend the software moat rather than the transistor: keep developers on the toolchain, sell whole racks instead of chips, and lock multi-year memory supply before rivals can
  2	Compute landlords	Structurally the weakest position: commodity capacity, 20-year debt against 3-year hardware, and hyperscaler customers who can insource. Power, not demand, caps growth	Sign long contracts and long power leases to convert a spot business into an annuity, and specialise on inference and sovereign hosting where hyperscalers are weak
  3	Frontier model labs	A rival’s open weights cap the price of everything they sell, and a four-point capability band removes the premium. Training cost still rises ~2.4×/yr	Move up into products and agents where the outcome can be priced, buy distribution outright, and go public to fund the next run
  4	Application and agent firms	Built on a supplier that can ship the same feature next quarter, and priced against a model layer that keeps getting cheaper	Own the workflow and the customer’s data rather than the model, sell the finished outcome, and stay multi-model so no lab can squeeze them
  5	Distribution platforms	They already have the users, so their risk is cannibalising a profitable business with a costlier one. Regulators are now prying their defaults open	Monetise attention through advertising rather than tokens, defend the default position, and comply early enough to keep it
  6	Data, tooling and evaluation	Every layer wants these functions in-house, and the labs ship free versions of the easy ones	Sell what a customer cannot build alone: proprietary data, independent evaluation, and audit that satisfies a regulator
  7	Embodied AI and robotics	Capability is no longer the constraint. Manufacturing scale, unit cost, safety and liability are, and China holds the manufacturing advantage	Buy scale early through partnerships and capital, target industrial work where liability is bounded, and treat the supply chain as the real product

  • Two pressures explain most of the migration between types. Everyone above the compute landlords is running from commoditisation, and everyone below the chipmakers is running from a supplier who might become a competitor. That is why labs are buying application firms and hyperscalers are designing silicon: the move is defensive in both directions.
  • The middle is the dangerous place to sit. Types 1 and 5 own something a rival cannot manufacture, a toolchain and a user base. Types 4 and 6 sit close enough to the customer to be paid for outcomes. Type 2 has neither, which is why a funding freeze prunes it first (A3).
  • Every type’s action list points at the same two assets. Whatever the row, the answer is a scarce input contracted long or a customer relationship held directly. Nothing else in this table survives a price war.

What This Means:

• Integration is now selective, not total. Owning a scarce layer still lowers marginal cost and locks in access. Owning an abundant one only ties up capital. That is why the same lab signs a 20-year power lease and rents accelerators from a competitor by the month.
• The split is by time horizon, not by layer. July 2026 saw the AI-natives integrate hard, on the argument that renting the whole stack stops working at scale. Both moves are running at once because they answer different questions: rent to stay flexible against a capability that changes every quarter, own to strip cost out of an inference workload that is now predictable. A firm that gets the two backwards pays twice.
• No integration is complete. Every player still depends on layers it doesn’t own (upstream silicon, non-AI rails, downstream specialist apps). The boundaries where stacks meet are A5’s subject.
• China’s super-app model is the most integrated stack in the world. Western players are catching up by integrating downward (compute, memory) and outward (protocols).
• The seven types compete for the same dollar, not with each other’s products. Every layer’s margin is another layer’s cost, so the fight over who keeps the token dollar runs vertically through the stack, not sideways between rivals (breakout A2-2).
• The types are migrating into each other, and the direction is defensive. Labs are buying application firms, hyperscalers are designing silicon, chipmakers are selling whole racks. Each move runs away from commoditisation above or from a supplier turning competitor below (breakout A2-3). Read a rival’s expansion as a statement about the pressure it is under, not about its ambition.
• Renting from a rival is a bet on the contract, not on the relationship. Anthropic’s capacity on a competitor’s data center is safe because it terminates on 90 days’ notice and the alternatives are real. Where a rented layer has no substitute and no exit, it is not a rental. It is a dependency priced as one, and it gets repriced the moment the landlord needs the capacity itself.
• Which type you are is the decision that sets the others. Capital intensity, margin, cycle exposure, and what can destroy you all follow from the row you occupy in breakout A2-2, not from your ambition. A lab that starts renting out surplus capacity has entered type 2 and inherited its debt-and-power risk, whatever its mission statement says.

Recommendations:

• Choose integration moves layer by layer, by scarcity. Own the 1–2 layers where scarcity and marginal-cost economics justify it. Rent the abundant ones, and keep the exit: Anthropic’s xAI capacity terminates on 90 days’ notice, which is what makes a rival’s data center safe to use.
• Watch the adjacent layer that will be contested first: chip-and-energy for hyperscalers, data-center for labs, agent-and-memory for platforms. Positioning in 2026–27 decides control through 2028.
• Treat layers you can’t own as managed dependencies, not commodities. Lock terms that secure capacity and data access. The contest to control those boundaries is A5’s.
• Price every rented layer against its substitute, and keep the substitute warm. A rented layer with a live alternative is a cost line. A rented layer without one is leverage held by someone else. Qualify a second supplier for each, even at a premium. The premium is what the exit clause costs.
• Name the one or two types you are, and refuse the others at the budget. Drifting into a second business model brings its capital needs and its failure mode with it. Decide deliberately whether you sell silicon, capacity, models, outcomes, attention, tooling, or machines.
• Apply one test before any move between types: does it end in a scarce input or a direct customer? Every type’s live strategy converges on those two assets (breakout A2-3). An expansion that reaches neither buys revenue that a price war removes, and adds a failure mode you did not have before.

↑ Back to Table of Contents

A3. Sustain the Capital Supercycle, Now Bound by Power

Summary: A capital supercycle funds the buildout. Sustaining it is the bet. Top-5 US tech capital spending is guided near $765B for 2026, about 3× 2024, with 2027 projected above $1T. The financing is increasingly circular, the clearest bubble signal, as NVIDIA’s commitment to OpenAI returns as chip orders. Spend runs near 90% of operating cash flow against a ~40% norm. July 2026 brought the first real pushback: Alphabet raised 2026 guidance again and its shares fell 7% the same day. The binding constraint has moved from chips to electricity, and interconnection queues now run to five years (breakout A3-1). Power has also become political, with more than 300 state bills in a single half-year. The private era is ending too, with Anthropic and OpenAI both filed to list. Price both the bubble and the payoff, reserve cash, and lock 20-year power now. A bust would hit smaller players hardest and concentrate the field rather than level it.

Evidence:

• Top-5 US tech capital spending: ~$235B (2024) → ~$400B (2025) → ~$750–765B guided for 2026, about ~3× 2024. The July 2026 earnings round raised it again: Alphabet lifted 2026 guidance from $180–190B to $195–205B on a second quarter of $44.9B, up ~100% year on year, Meta moved to $125–145B, and Amazon guided about $200B (CNBC). That is a third consecutive year of 60%+ growth, with analysts higher still (Morgan Stanley ~$805B) and 2027 projected above $1T (Goldman Sachs, CNBC). The labs’ own commitments sit on top of that: OpenAI’s Stargate ($500B 2025–2029), Anthropic (>$100B to AWS + ~1M Google chips), xAI ($20B raised). These turn into balance-sheet exposure at the neoclouds, where Oracle now carries a $638B backlog, >50% of it OpenAI (Oracle). Counting all data centers, Dell’Oro puts 2026 capital spending above $1T (Dell’Oro).
• AI capital spending now drives an outsized share of US GDP growth. The exact share turns on the window and what is counted, and it is rising. Two careful cuts bracket it. Data-center investment alone contributed ~40% of real GDP growth across Q1–Q3 2025 (St. Louis Fed). For the single quarter Q1 2026, adding software and IP products lifts it to ~74% (≈1.55 of 2.1 points, analyst-derived from the BEA third estimate, TFTC/BEA). The gap is window (a three-quarter average vs. one quarter) and base (equipment only vs. equipment-plus-software). The trend is up quarter on quarter, the signal that matters more than any single figure. As a share of output rather than growth, AI capital spending reached ~2.4% of GDP and is projected to peak near ~3%, approaching the 1850s railway peak and past the dot-com telecom peak (Fortune).
• The spend runs near ~90% of operating cash flow (BofA estimates ~94% in 2026, vs a ~40% norm), AI-related debt issuance is set to top $500B in 2026 (BIS), and Bain estimates AI needs ~$2T in new annual revenue by 2030, on course to fall ~$800B short (Bain). Debt, not just equity, now carries the buildout, and it is concentrated in the thinnest balance sheets. The five largest cloud firms raised ~$108B of bonds in 2025 and roughly $100B more in early 2026, with Morgan Stanley projecting $250–300B of hyperscaler issuance for 2026. The neoclouds are more exposed: CoreWeave carries ~$34.7B of long-term debt and Nebius ~$9.5B against far smaller revenue (io-fund). The financing is increasingly circular, the cycle’s clearest bubble signal. NVIDIA’s up-to-$100B commitment to OpenAI returns as chip orders routed through Oracle and the neoclouds, and similar chip-for-equity and compute-for-revenue loops link the largest players (Bloomberg). The NVIDIA–OpenAI tranche reportedly stalled in February 2026 over OpenAI’s “lack of financial discipline.” When the same dollars count as revenue at one node and spend at the next, demand looks firmer than it is.
• China runs a parallel supercycle, state-directed rather than market-funded. Beijing prepared a ~RMB 2T (~$295B) five-year plan for a national AI data-center network, operated by China Mobile and China Telecom with at least 80% domestic core technology and a target around 2028 (Bloomberg). Total Chinese AI investment reached ~RMB 890B (~$125B) in 2026, roughly 38% of the global figure, with government money the largest single share at ~39%. Private programmes run on top (Alibaba RMB 380B / ~$54B over three years), and AI took ~80% of a record ~$300B Q1-2026 venture-capital quarter (Crunchbase).
• Every large AI firm has now bought nuclear, and the public has started to push back. Microsoft, Google, Amazon and Meta each signed at least one nuclear agreement in 2026, more than a dozen deals totalling close to 10 GW, with Meta alone contracting up to 6.6 GW (smrintel). Against that, the political weather turned: Illinois suspended state tax incentives for new data centers for two years from 1 July 2026, and lawmakers in more than 30 states introduced over 300 data-center bills in the first half of 2026. Cheap power and a welcome are no longer the same thing. Physical chokepoints bind the spend: data-center electricity demand +17% in 2025, small modular reactor offtakes 25→45 GW, and ~$156B of projects blocked or delayed in 2025 (IEA, Data Center Watch).
• The market pushed back for the first time. Alphabet’s shares fell ~7% on the day it raised capital spending guidance, and Amazon, Meta and Microsoft fell with it, on scrutiny of infrastructure spending against shrinking cash and unproven returns (CNBC). For three years a capital spending raise read as confidence. In July 2026 it read as risk. That is the investor-patience limit this chapter prices, arriving early. Counter-signals: MIT (Aug 2025) found 95% of GenAI investment with zero measurable return (Fortune) and NBER (Feb 2026) ~89% of firms reporting no productivity impact (NBER), set against Google’s AI backlog doubling to $460B in one quarter (the pilot-to-production signal). The bubble-or-delayed-payoff question peaks in 2026–27: either pilot ROI visibly improves by ~2027 or valuations face structural pressure. A bust would hit smaller, less-capitalized players far harder than the cash-rich giants, so a correction concentrates the field rather than levelling it (the consolidation route into W3).
• The private era is ending: both leading labs have filed to go public. Anthropic filed confidentially on 1 June 2026 and targets an October Nasdaq listing, which would be the first debut near a $1T valuation. OpenAI filed on 8 June and is leaning toward 2027 after market volatility (Forbes). Listing converts a mission-governed private lab into a company with quarterly disclosure, an index-tracking shareholder base, and a share price that funds acquisitions.

• Breakout A3-1: the constraint moved from chips to electricity, and the capital is borrowed against it.

  Scarcity has migrated down the stack: intelligence, then accelerators, then memory, now power. Each move was bought by spending on the next input down, and each next input takes longer to add, so a firm can now hold chips it cannot switch on. The table sets each link of the capital chain against what actually binds it, roughly in the order the constraint travelled.

  Link	2026 scale	What binds it	Lead time to relieve
  Capital committed	~$765B top-5 US guided spend, ~3× 2024; >$1T projected 2027	Investor patience and the 2027 return test	One earnings cycle
  Debt raised against it	~$108B (2025) plus ~$100B early 2026; $250–300B projected for 2026	Interest cover if revenue lags; neoclouds most exposed	Refinancing window
  Chips	~$281B of intelligent data-center silicon	Advanced packaging booked into 2027; memory sold out	18–24 months
  Electricity	~20–30 GW of AI data-center demand by late 2027, doubling by 2027	Interconnection queues reported near 2,600 GW, with waits reaching five years	3–7 years
  Grid equipment and sites	~11 GW announced for 2026 not yet under construction	Transformers, turbines, permits, and now legislatures: 300+ state bills in H1 2026, Illinois incentives suspended from 1 July	2–5 years
  Realized capacity	About half of global projects delayed	The slowest link above, not the fastest	Set by power, not chips
  China, in parallel	~$295B state plan to 2028, plus ~$125B of total 2026 AI investment, ~39% government	Domestic chips and memory, not capital or permits	State-set, and it runs through a Western funding freeze

  • The order of the constraints inverted. A chip order can be filled in quarters. A grid connection takes years. So the marginal gigawatt, not the marginal accelerator, now decides who trains the next frontier model.
  • This is why energy-rich states entered the race. The Gulf pairs surplus gas and solar with sovereign capital, which is why the US–UAE partnership frames a 5 GW campus and Stargate UAE a 1 GW facility. Their advantage is not talent or chips. It is interconnection they can approve themselves.
  • Borrowed money against a slow physical asset is the fragility. Debt is priced on schedules that assume power arrives. When it slips, interest runs anyway. That is the mechanism by which a delay, not a demand collapse, would trigger the correction.

What This Means:

• The self-reinforcing loop (spend, compute, capability, revenue, more spend) dominates through 2027–28. AI capability moves in months, adoption in years, productivity in decades. Those timescales are pulling apart, and neither markets nor politics will wait the decade it takes for the payoff to arrive.
• This rhymes with past build-outs, but the two stacks are exposed differently. Railways, electrification, and the dot-com fiber boom all over-built, busted, and left durable infrastructure behind, so a correction is normal, not the end of the technology. The difference is who carries the risk. The US supercycle is private, equity- and debt-funded, exposed to investor patience and a market repricing, while China’s is state-backed and steered. A 2026–27 bust hits the US field harder and concentrates it, while China’s build-out runs through the dip.
• Profit concentration is unprecedented: NVIDIA earns above 70% gross margins at the upstream bottleneck, and it is arguably without precedent for GDP growth this high to be driven by profits captured by so few firms (Bridgewater).
• Physical chokepoints (advanced packaging, memory, grid power, permits, opposition) keep deployable compute below paper capacity through 2027. Supply is the binding constraint, not demand.
• The scarce thing keeps moving down the stack, and each step takes longer to fix than the last. Intelligence was scarce until the frontier converged into a four-point band (A1). Then accelerators, until custom silicon and a second supply curve eased it (A2, A4). Then memory, now sold out with demand exceeding capacity past 2030. Now power, with interconnection queues reaching five years. The sequence is not coincidence: each bottleneck was relieved by pouring capital into the next input down, and each next input has a longer lead time, so every round of relief buys less time than the one before. Intelligence → accelerators → memory → power, and the link after power has no supplier at all: political consent to build (breakout A3-1).
• A capital plan built around chip allocation is solving last year’s problem. Chips arrive in quarters, memory in years, grid connections in half a decade. Plan against the slowest link you depend on, not the one your vendor talks about.

Recommendations:

• Plan capital allocation against two 2027 scenarios: pilot ROI visibly improves and validates today’s capital spending, or it doesn’t and valuations correct sharply. Price both. Reserve enough cash to survive the correction without forced asset sales.
• Lock in 20-year power, water, and land now, while small modular reactor terms hold. Offtakes grew 25→45 GW in 18 months and the favorable-terms window narrows each quarter.
• Build community engagement into data-center projects from day one, because the opposition is now legislative. ~$156B was blocked or delayed in 2025, and in the first half of 2026 more than 30 states introduced over 300 data-center bills, with Illinois suspending incentives outright. Treat social license as a core function, not PR, and assume the siting rules you plan against can change mid-build.
• Diversify silicon beyond NVIDIA (AMD, Google Google’s own AI chips, AWS Trainium, custom chips) and build workload portability so capacity can shift as supply tightens.
• Price for compute scarcity rather than absorbing margin compression: API increases, multi-year premiums, dedicated-capacity charges. This pricing power narrows once supply catches up.

↑ Back to Table of Contents

A4. Race the Frontier in a Two-Stack World

Summary: The frontier is a two-stack race. No single bottleneck can stall it. The US holds the capability lead with premium closed models. China ships at the frontier on a 4–6 week cadence and prices 5–30× cheaper, with GLM-5.2 topping its harder professional version ahead of GPT-5.5. Performance commoditizes in 6–12 months under open-weight pressure. Self-improvement compounds it: Anthropic reports AI is involved in over 80% of code merged inside the company. The biggest non-technical hazard is a lab’s own state: in June 2026 a US export-control order took Anthropic’s newest models off the market worldwide for 19 days. The cheap-China assumption is also weakening, with open-weight prices rising sharply in one generation. Decide now whether you will be one of the 5–7 frontier firms by 2028.

Evidence:

• Multi-axis scaling: reasoning-training compute grows 10× every 3–5 months and converges with pretraining by 2026 (Epoch AI). The US–China gap on broad knowledge tests narrowed from 17.5 to 0.3 points in a year and the open–closed gap (LMArena) from 8.0% to 1.7% (Stanford HAI). Quality public text nears exhaustion (2026–2032, Epoch 80% CI), so synthetic and AI-filtered data fill an ever-larger share of training inputs (Epoch). The data path stays open.
• Frontier agents sustain hours of autonomous work. METR’s mid-2026 top measurement is ~11 h, with wide error bars as the suite saturates, the 50%-task horizon doubling every ~89 days (METR). Naive extrapolation reaches work-week autonomy around 2027–28.
• Seven Chinese labs ship at the frontier on a 4–6 week cadence (DeepSeek, Qwen, Z.ai, Moonshot, MiniMax, ByteDance, Tencent). The open-weight GLM-5.2 (MIT license) tops its harder professional version ahead of GPT-5.5 (VentureBeat) and Kimi K2.6 leads the open-weight field on agentic browsing. Z.ai’s GLM-5 was trained entirely on Huawei (non-US) hardware (SCMP). The parameter race continues in the open: Moonshot released Kimi K3 at 2.8 trillion parameters on 17 July 2026, and Alibaba answered with Qwen 3.8 at 2.4 trillion, both open-weight (Cybernews).
• China has now won the usage share, not only the price argument. By May 2026 Chinese-origin models were ~61% of all tokens routed through OpenRouter, the largest neutral model router, and four of the five most-used models were Chinese. DeepSeek alone ran ~17.6% against Anthropic’s 15.4%, Google fell from ~37% to ~13%, and Meta’s Llama dropped below 1%, off the rankings (Data Gravity). Total routed volume quadrupled year on year, so the cheap models expanded the market rather than only taking share.
• A government took a frontier model off the world market for 19 days. A US export-control order of 12 June 2026 required Anthropic to bar foreign nationals from Claude Fable 5 and Mythos 5, inside and outside the US and including its own foreign-national staff, over concerns that safeguards on vulnerability-finding could be jailbroken. Anthropic suspended global access to both. The administration lifted the order on 30 June and the models returned on 1 July with an added cybersecurity classifier (CNBC, VentureBeat). Enterprise customers lost their supplier for nearly three weeks with no commercial recourse.
• The cheap-China assumption is weakening. Moonshot priced Kimi K3 at $3.00 per million input tokens against $0.95 for K2.6, and $15.00 output against $4.00, roughly tripling and quadrupling in one generation (The Decoder). DeepSeek announced peak-hour pricing for V4 from 30 June 2026, doubling rates in two daily windows. Per finished task the gap is now narrow: K3 averages ~$0.94 against GPT-5.6 Sol’s ~$1.04. DeepSeek V4 Pro runs ~15–30× cheaper blended, and ~34× on output tokens, than GPT-5.5/Claude Opus 4.8. Chinese frontier pricing runs 5–30× cheaper broadly (Artificial Analysis).
• Two cost curves move in opposite directions. Pushing the frontier costs ~2.4×/yr (Epoch crosses ~$1B/run by 2027, labs project ~$10B by 2028), while reaching an existing capability falls ~10×/yr. A GPT-3.5-level system got >280× cheaper in ~18 months (Stanford HAI). China rides the second curve: DeepSeek-V3’s cited ~$5.6M is only the final run, against a fleet SemiAnalysis values at ~$1.6B. Cheap Chinese models and $10B frontier runs are two ends of one cost structure, not a contradiction. US inference is already gross-margin positive, so the labs’ losses sit on the frontier-and-capacity line, not the token price (AI After Hours). Rising per-run cost concentrates the field to ~5–7 firms and sovereigns by 2028: Anthropic, OpenAI, Google, xAI, and Meta in the US, DeepSeek and Qwen in China, plus sovereign-backed efforts.

• Breakout A4-1: AI improving AI is the engine under the frontier race, and its doubling times are measured in months.

  The frontier race is increasingly run by AI on AI, the fastest force behind every timeline in this lens. Once a capability can train itself, it shoots past human level fast, and that loop is now closing on AI development itself. The tell-tale signal is that the load-bearing quantities double in months, not years.

  Domain	How AI improves AI	Realized impact	Forecast	Reference
  Self-play (proof of principle, 2017)	Learns from self-play alone, no human data. The reward is the game result.	AlphaGo Zero beat the Lee-Sedol-beating version 100–0 in 3 days, zero human games. AlphaZero then generalized to chess and shogi.	The template, self-made data plus a verifiable reward, now underpins reasoning models (last row).	Nature 2017
  Autonomous AI R&D (2024–)	Agents run real ML-research tasks end to end. Capability tracked over time.	Task-completion horizondoubles ~every 3–4 months since 2024, from ~7 months before. Top agents scored 4× human experts at a 2-hour budget.	Work-week autonomy ~2027–28. AI-2027 lead author expects AI research fully automated ~2029 (AI 2027).	METR; RE-Bench
  Software / coding (2025–26)	Frontier models write, review, and merge the code that builds frontier models.	Anthropic reports AI is involved in >80% of merged code inside the company, up from low single digits in early 2025. Google’s new code passed ~75% AI-written by April 2026.	Trends toward near-100% of routine code.	Anthropic; Alphabet Q1 2026
  Algorithm & kernel design (2025)	A Gemini agent evolves code against automatic evaluators, keeping what measurably wins.	AlphaEvolve recovered ~0.7% of Google’s fleet compute, sped a Gemini training kernel +23% (~1% off training time), and FlashAttention 32.5%.	DeepMind aims it recursively at Gemini’s own training stack.	DeepMind
  Reasoning via RL, then distillation (2025)	Rewarded only for correct answers, the model generates its own reasoning traces, then teaches smaller models.	DeepSeek-R1-Zero, pure RL with no human reasoning labels, lifted AIME 2024 from 15.6% → 77.9%. Distilled into a 3.8B model at 94.6% on Math-500.	The dominant 2025–26 recipe. Open small models close the gap in 6–12 months.	Nature 2025; Phi-4-mini

• Breakout A4-3: two cost curves move in opposite directions, and both are true at once.

  The most confusing number in AI is the cost of a model, because two different costs are quoted as one. Pushing the frontier gets more expensive every year. Reaching a capability that already exists gets radically cheaper every year. A $10B training run and a model 30× cheaper than last year’s are not a contradiction. They sit at opposite ends of the same structure, and which curve a firm rides decides its whole strategy.

  	Cost to push the frontier	Cost to reach an existing capability
  Direction	Rising ~2.4×/yr	Falling ~10×/yr
  Evidence	Epoch crosses ~$1B per run by 2027; labs project ~$10B by 2028	A GPT-3.5-level system got >280× cheaper in ~18 months (Stanford HAI)
  What drives it	More compute, longer reasoning training, scarcer power and packaging	Distillation, better hardware, caching, open weights following months behind
  Who rides it	The 5–7 firms and sovereigns who can still fund a run	Everyone else, including every Chinese open-weight lab
  Strategic consequence	Entry closes; the field concentrates by arithmetic, not by merger	Any given capability becomes a commodity within 6–12 months of arriving
  The trap	Funding a run you cannot amortize before the capability commoditizes	Building a business on a price advantage that your supplier can withdraw

  • Both curves are the same fact seen from two ends. Yesterday’s frontier is today’s commodity. The gap between the curves is the window in which a frontier firm has to earn back the run, and that window is 6–12 months and shrinking.
  • This is why cheap Chinese pricing is not evidence of cheap Chinese training. DeepSeek-V3’s cited ~$5.6M is the final run only, against a fleet SemiAnalysis values near ~$1.6B. China rides the falling curve deliberately and prices against it.
  • Pick which curve your business sits on, because you cannot sit on both. Riding the rising curve requires capital that only concentrates further. Riding the falling curve requires accepting that your differentiation is never the model itself. The firms in trouble are the ones that funded the first and priced as though they were on the second.

• Breakout A4-4: two ways to build an AI industry — America stacks specialists, China stacks one company.

  Both systems build the same seven layers (A2). They differ in who owns the seams between them. In the US each layer is a separate company selling to the next, so every boundary is a market with a price and a contract. In China the layers sit inside one conglomerate or one state programme, so the boundaries are internal decisions. The comparison below is organised by what each row does to speed, cost, and the ability to switch — the three things a buyer actually feels.

  Dimension	US — layered specialists	China — integrated silos	What the difference does
  Who owns a layer	A different company per layer, each selling outward	One conglomerate or state programme spans chips to app	US buyers can mix; Chinese users get one bundle
  Where the boundary sits	A market: priced, contracted, contestable	Internal: a transfer price, not a negotiation	US seams are where startups enter and antitrust acts
  Who funds it	Private capital, ~$750B+ in 2026, exposed to investor patience	State-directed, ~$295B five-year plan plus private capex, ~39% government	A market correction prunes the US field and skips China’s
  How it is steered	Competition; policy acts afterwards and from outside	Planning; targets set the direction, e.g. 80% domestic core technology	China aims capacity, the US discovers it
  Model posture	Premium closed models, sold by the token	Open weights as the outward-facing interface, ~61% of routed tokens	Open release is China’s export channel and a cap on US rent
  Silicon	NVIDIA plus hyperscaler custom parts, best-in-class	Huawei Ascend and domestic parts, ~41% of China’s market, capped by memory	The US leads on parts, China on not needing anyone’s permission
  Deployment speed	Slower to assemble: five firms must agree	Faster: rails pre-bundled inside one login and one wallet	China reaches scale first, the US reaches quality first
  What can break it	A funding freeze, or a supplier turning competitor	A supply cut-off in memory or tooling, or a policy reversal	Different failure modes, so both can be true at once

  • The commonalities are larger than the contrast usually admits. Both run the same seven layers, both concentrate to a handful of players, both hit the same physical limits in memory and power, and both are now moving toward the middle: US labs are integrating downward while Chinese frontier prices rise toward Western levels (A2, A4). The systems differ in ownership of the seams, not in physics or economics.
  • Contestability is the real dividing line, and it cuts both ways. A US seam is a place a startup can enter, a regulator can pry open, and a customer can switch at. It is also a place where five firms must agree before anything ships. Chinese integration removes the friction and the entry point together.
  • Each system’s strength is the other’s exposure. Private funding gives the US speed and a bust; state funding gives China continuity and misallocation. Closed premium models earn rent until open weights cap it; open weights win distribution until someone has to pay for the next training run.
  • For a Western firm the practical read is narrow. Do not plan to out-integrate a Chinese conglomerate on its own ground. Plan to be the layer others must buy, and to be switchable enough that a buyer choosing you is not choosing a decade.

What This Means:

• No single bottleneck can stall capability. Three parallel paths are the technical reason CEO timelines got shorter in 2025, not longer.
• Performance commoditizes in 6–12 months under Chinese open-weight pressure, faster than any prior tech cycle. Chinese labs now innovate in directions the US does not (Kimi’s 100+ parallel sub-agents, GLM-5 on a homegrown chip stack).
• US export controls became an accelerator, not a brake. They sped China’s drive to build its own stack rather than slowing it.
• Government interference is now a frontier-race risk of its own, and it escalated in 2026. A lab’s biggest non-technical hazard is its own state. First the US labeled Anthropic a “supply-chain risk” and ordered agencies to stop using it after it refused a Pentagon “all lawful purposes” demand, until a court blocked it as illegal retaliation (G3). Then, in June, an export-control order removed its newest models from the global market for 19 days. The escalation matters: the first action cut a supplier off from public demand, the second cut it off from every customer it had. A frontier model is now treated as a controlled export, so a lab’s product can be switched off by its own government on a security judgement it cannot appeal in time. Capital does not price this.
• The measured capabilities of US and Chinese labs are closer than debate suggests. Differentiation is moving from saturating benchmarks to deployed capability and ecosystem.
• Is the price gap sustainable, and open weights stable? The gap is partly structural (China rides the falling cost-to-reach curve) and partly subsidized (cheap state compute, loss-leader pricing), so expect a durable discount, not near-zero pricing. 2026 showed the narrowing: the newest Chinese frontier model tripled its input price and quadrupled its output price in one generation, and per finished task now costs about the same as its Western rival. Cheap was a phase of the land grab, not a property of the stack. Open-weight release is a strategy, not a guarantee. It persists while it serves China’s interest in commoditizing the US lead, and is the swing factor for W5.

Recommendations:

• Fund all three capability paths in parallel (pretraining, post-training, especially reinforcement learning with verifiable rewards, and inference-time compute) so no single bottleneck can stall you.
• Forecast compute against the agent task horizon, not benchmark scores. Project from today’s ~11-hour horizon along METR’s ~89-day doubling.
• Replace saturated benchmarks with measures of economic value in internal evaluation: expert judgment on real tasks, unsupervised work-time, return on use cases (OpenAI’s economic-value benchmark is one example).
• Decide now whether you will be one of the 5–7 frontier firms by 2028, or not. If yes, invest in operational strength (evaluation, monitoring, governance). If not, say so and build depth in a vertical app or tooling instead.
• Choose your business model deliberately: premium closed models defended by operational strength, or open weights at a structurally lower price. A hybrid must be chosen on purpose. Watch Chinese frontier pricing as the forward floor your own pricing must meet within 6–12 months.

↑ Back to Table of Contents

A5. Capture the Customer, Then Convert It Into Revenue

Summary: Competing for the customer successfully requires getting their attention and recording their data to build a bond. An agent’s memory of the customer is the main source of defensibility. Portable memory weakens it. The agent-protocol layer is the next platform war, fought across tool-access, payment, and identity standards. An interface pays rent only where it is also a chokepoint, which is why the tool-access standard is open and Visa is not. Capture is not yet revenue. 2026 resolved the money question into four models, and advertising arrived: ChatGPT ads passed $100M annualized within two months of launch (breakout A5-1). The scale is now large and the two leaders are converging: by July 2026 Anthropic ran at ~$74B and OpenAI at ~$41B annualized, and OpenAI’s enterprise share passed 40% on its way to parity with consumer. Realized prices sit far below sticker, because agentic traffic caches heavily. Treat user data and agent memory as Day 1 priorities, and choose a revenue model on purpose. Firms 12–18 months behind stay behind.

Evidence:

• The revenue is now very large, and the two business models are converging. By July 2026 Anthropic reported roughly $74B annualized against OpenAI’s ~$41B, having grown 58% and 65% respectively. Anthropic leads the enterprise model market at ~32% API share to OpenAI’s ~25%, and Claude Code holds ~54% of AI coding spend. OpenAI, long the consumer company, now takes over 40% of revenue from enterprise products and expects parity with consumer by the end of 2026, credited to the GPT-5.6 series, the ChatGPT Work agent, and Codex (CNBC). Outcome-based pricing replaces seat-based software subscriptions by 2027–28. Per-seat pricing breaks once agents do the work of multiple seats.
• Supporting several agent protocols measurably pays. Merchants implementing more than one agentic-commerce standard see roughly 40% more agent-driven traffic than single-protocol adopters, and AI referral traffic to US retail sites grew ~393% year on year in the first quarter of 2026. OpenAI meanwhile shut its Instant Checkout initiative in March 2026 on weak performance and refocused on discovery and in-chat apps.
• The application layer now outgrows the labs it runs on. Cursor reached ~$4B annualized in under four years, ~$2.6B of it enterprise, and SpaceX bought its parent Anysphere for $60B in stock on 16 June 2026, the largest startup acquisition on record (CNBC). The buyer wanted a coding engine and a developer channel for Grok, not a model. Vertical AI revenue: $3.5B in 2025 (~3× 2024). 10 products generate $1B+ annual recurring revenue, 50 over $100M (Menlo Ventures). Legal leads (Harvey ~$300M annual recurring revenue/$11B valuation, Clio $500M annual recurring revenue), detail in B2.
• Agents are shipping, not piloting. ~80% of enterprise applications shipped or updated in Q1 2026 embed at least one AI agent, up from ~33% in 2024, and the median deployment reaches measurable value in ~5.1 months. IDC projects a 10× rise in agent usage and ~1000× growth in inference demand by 2027.
• Advertising arrived as the consumer revenue model. OpenAI launched ads in ChatGPT in February 2026 for free and low-tier users, passed $100M annualized within two months, and targets ~$2.5B for 2026 and ~$25B by 2028 at roughly $60 per thousand impressions, about 3× Meta’s rate (IntuitionLabs). Google is bringing ads to Gemini. Perplexity tested and abandoned ads on trust grounds, moving to subscriptions instead.
• Realized token prices sit far below sticker, which is why volume growth and price collapse coexist. Blended realized price for one frontier model ran ~$0.99 per million tokens against a $5/$25 published rate, because agentic workloads run roughly 300:1 input to output with cache-hit rates above 90%. Anthropic’s inference gross margin rose from 38% to over 70% across 2026 on that mix (SemiAnalysis).
• An agent platform’s memory of the user becomes the main source of defensibility in consumer markets through 2026–27. By 2028 the defensible position is proprietary data + memory + integrations. Memory is also being commoditized. In early 2026 ChatGPT, Claude, and Gemini all moved memory to free tiers and Claude added one-click import of a rival’s memory (9to5Mac). Portable memory weakens the lock-in it creates, the swing factor for whether the moat holds (I2). Hyper-personalization market ~$21.8B (2024), ~18% annual growth. Firms that excel at AI personalization earn ~40% more revenue from those activities (McKinsey). Non-AI rails are unavoidable and stable in the West. Payments (Visa, Stripe), identity (Apple/Google ID), logistics (Amazon, FedEx) sit outside any AI stack. In China the same rails sit inside the closed ecosystem (Alipay, WeChat Pay owned by the conglomerates), so there is no external seam to contest.

• Breakout A5-1: four ways an AI company gets paid, and each one buys a different kind of customer.

  Capturing attention is not the same as booking revenue. By 2026 four models had separated out, and the choice determines who the customer is, what margin looks like, and what can take it away. Most large players run more than one, but each has a dominant model that shapes everything else.

  Model	Who pays, and for what	2026 evidence	Margin behavior	What kills it
  Consumer subscription	A person pays monthly for better access and higher limits	Long ~85% of OpenAI’s revenue, now under 60% as enterprise passes 40% and heads for parity; Perplexity targets ~$500M of subscription revenue after dropping ads	Improves as inference costs fall; churn-sensitive	A free rival that is good enough
  Usage and tokens	A developer or firm pays per unit of work the model does	Realized ~$0.99/M tokens against $5/$25 sticker, on ~300:1 input-output and >90% cache hits; inference margin 38% → >70%	Rises with caching and hardware, falls with price war	Open weights at 3–12× lower price
  Advertising	A merchant pays for placement in front of the user’s question	ChatGPT ads passed $100M annualized in two months, targeting ~$2.5B (2026) and ~$25B (2028) at ~$60 per thousand impressions	Highest, and scales with attention not compute	User trust, and regulation of AI answers
  Enterprise licence and outcomes	A company pays for a job completed inside its workflow	~85% of Anthropic’s revenue is business, at ~$74B annualized; ~32% of the enterprise model market against OpenAI’s ~25%, and ~54% of coding spend; Cursor ~$2.6B enterprise	Stable and contracted; slowest to win	The buyer’s own build, or the platform shipping the feature

  • The two leading labs started as opposite companies and are converging. Anthropic still earns roughly 85% from business. OpenAI was ~85% consumer and is now past 40% enterprise, heading for parity within the year. The original split explains the divergent moves of 2026, one launching advertising and the other a partner network. The convergence explains what comes next: both will run two revenue models at once, and each will meet the other’s failure mode as well as its own.
  • Advertising is the model with the best economics and the worst alignment. It pays per unit of attention, not per unit of compute, so it escapes the token price war entirely. It also gives the seller a reason to shape an answer, which is the conflict regulators will reach first.
  • Falling prices are not falling revenue. Published prices fell around 1000× in three years, yet lab revenue and margin both rose, because volume and caching grew faster than price fell. Anyone reading the price collapse as a margin collapse has the sign wrong.

What This Means:

• Why attention is the whole game. Everything here reduces to one scarce input: the user’s attention and the data it throws off. Attention converts into profit. It sells advertising, produces the behavioral data that trains and personalizes models, and holds a user inside one agent long enough that leaving gets costly. The agent economy monetizes it through memory and the fees the platform takes on what the agent does. The customer fight turns on attention and memory above all, though never only them: intelligence, loyalty, price, and speed each win share, and the durable advantage is the bundle. The same attention is the individual’s keystone in Lens I (I3).
• The user-data advantage is new and structural. It compounds with every interaction, and the old cookie moat has moved up one layer to the agent platform, so merchants lose their memory of the customer by default unless they build their own agent integrations.
• The agent-protocol layer is the next platform competition, fought on three fronts at once (most users, winning standard, which stack the specialist apps plug into). A stack can win one and lose the others.
• An interface pays rent only where it is also a chokepoint. An open standard buys ecosystem position and a view of the layer above, not a toll. The toll arrives only when the standard is welded to a scarce owned resource, which is why the tool-access standard (MCP) is open and Visa is not. The whole contest is a property of open stacks: China’s seams are internalized or state-mandated, so contestability favors the West (B4).
• Deep personalization differentiates the product itself, not just the marketing. Firms that treat it as a tactic miss the opportunity.
• The revenue model is a strategic choice, not a billing detail (breakout A5-1). It decides who the customer is, which costs matter, and what can destroy the business. A consumer-subscription firm fears a free rival. A token business fears open weights. An advertising business fears losing trust. Running two models well is possible. Drifting between them is not.
• Value capture moved up the stack in 2026. Through 2023–25 the infrastructure layer took almost all of it. This year the labs took the larger share, on rising volume and better inference economics, while foundry and chip pricing stayed comparatively restrained (SemiAnalysis). Where profit sits is not fixed, and it has moved once already.

Recommendations:

• Choose the revenue model deliberately, and staff for its failure mode. Pick the dominant model from breakout A5-1 and build the defense it needs: cost per query for subscriptions, caching and routing for tokens, trust and disclosure for advertising, and evaluation and integration for enterprise outcomes.
• Treat user data and agent memory as Day 1 priorities. Memory built across interactions becomes a structural advantage by 2028. Firms 12–18 months behind stay behind.
• Position early in agent-protocol standard-setting, but know what it buys. Supporting several protocols early keeps options open. Setting a standard buys ecosystem position and a view of the layer above, not a toll. Own the chokepoint it rides on (a settlement rail, scarce capacity) too.
• Build interfaces to payments, identity, and logistics as partnerships, not bolt-ons. These rails aren’t getting absorbed (in China they sit inside the incumbent’s closed ecosystem). The play is terms that lock in cross-stack data and revenue share.
• Invest in operational AI (evaluation, monitoring, governance, provenance). Capital underprices the layer that decides what you can actually deploy as catastrophe-capable chemical, biological, radiological and nuclear thresholds and EU AI Act enforcement arrive.
• Move enterprise contracts from per-seat to outcome-based pricing in 2026 renewals. Per-seat breaks once agents do the work of multiple seats. Restructure while compute scarcity still gives you leverage.

↑ Back to Table of Contents

A6. Turn an Understanding of the Other Lenses Into AI-Company Advantage

Summary: The first five chapters build Lens A from the inside — the stack, the capital, the frontier race, the customer. This one reads the rest of the board for the same end: advantage. Each other lens is a force that lands on your growth and your margin. The businesses that must deploy what you build (B), the people you hire, sell to, and displace (I), the governments that fund you, buy from you, and can cut you off (G), and the system that decides whether your rents survive (W). Read for your own interest, each yields a move the build-and-race chapters miss. Anticipate the shift, act while it is still optional, and turn someone else’s disruption into demand for what you sell.

Lens B — businesses. Businesses race to deploy AI before a rival reprices their sector, yet ~95% of pilots fail and only ~1% call their strategy mature. The block is organizational, not the model (B3). They pay for outcomes, not tokens, and competition passes most of AI’s cost reduction through to their customers as lower prices (B5). For a supplier, the buyer’s real bottleneck is implementation, and the durable spend follows whoever closes it. The market has already voted: the lab that leads enterprise share also leads coding spend, and the consumer-first lab is racing to enterprise parity within the year (A5). Derived recommendations:

• Sell the implementation gap, not the model. ~95% of enterprise pilots show no P&L in six months, and the failures are organizational — wrong project, no baseline, unready data (MIT/Fortune). What the buyer lacks is readiness, not raw capability. Package evaluation, workflow redesign, and governance as the product, and price on the profit you unlock, not the tokens you meter.
• Bank the outcome-pricing shift before rivals force it on you. Buyers move from per-seat to outcome-based pricing as agents do the work of many seats, and ~53% already buy rather than build (Menlo Ventures). Restructure enterprise contracts to charge for resolved tickets, closed deals, or booked savings while compute scarcity still gives you pricing power (A3).
• Arm the buyer’s moat, because that is where your revenue sticks. For a non-AI firm the surviving moats are proprietary data and customer-and-agent memory (B5-2). Sell the tooling that builds them — private data pipelines, memory, orchestration on the customer’s own data — so your platform becomes the layer their defensibility depends on, not a model they can swap.

Lens I — individuals. Individuals meet AI as workers, consumers, and citizens, each acting in self-interest. AI fluency pays a 62% wage premium while the entry ladder shortens, unguarded reliance erodes the skill it appears to build, and the shared-trust layer falls as deepfakes scale ~16× (I1, I3, I5). For an AI firm, those individuals are your scarce talent, your consumer market, and the source of a trust backlash that can hit your license before regulation does. Derived recommendations:

• Hire from the churn you create. AI fluency commands a 62% premium and keeps widening (PwC), while entry-level and exposed workers are cut first (22–25-year-olds −13%, Stanford). The same displacement that grows your market floods the labor pool with retrainable experienced talent. Recruit the AI-fluent early, and convert displaced specialists into the evaluation, red-team, and domain-tuning roles you cannot otherwise fill.
• Design for skill-building, not dependency — it is cheaper than the regulation that follows harm. Unguarded tutors cut exam scores 17% and heavy companion use tracks with loneliness and dependence (PNAS), and regulators have already moved (Illinois, Nevada, and Utah restricted AI therapy; the FTC opened a companion-chatbot inquiry, FTC). Build scaffolding that makes users think, and safety rails on companion products, ahead of the rules — a product edge and a pre-emption at once.
• Ship provenance before the trust collapse taxes every model’s output. Deepfakes rose ~16× toward ~8M and the default flipped to “probably fake” (Sumsub). A polluted information layer erodes trust in AI-generated content generally, your own included. Adopt content credentials across your outputs, so authenticated media is a feature you sell, not a liability you carry.

Lens G — governments. Governments try to govern at AI’s pace from bounded self-interest — competition abroad, legitimacy at home (G1). They are moving catastrophe-capable AI under licensing, buying AI at scale ($200M Pentagon contracts each), and have shown they will cut a supplier off. The US barred Anthropic from Defense systems over a usage-policy refusal, then took its newest models off the world market for 19 days under export control (A4, G3, G4). From 2 August 2026 the EU can demand access to a general-purpose model and order it off the market. For an AI firm, government is at once your largest anchor customer, your rule-setter, and a single point of political failure. Derived recommendations:

• Treat your own state as a frontier-race risk, and diversify across jurisdictions. A lab’s biggest non-technical hazard is its own government. Anthropic was labeled a “supply-chain risk” and barred from Defense systems after refusing an “all lawful purposes” demand (CNBC). Spread public-sector revenue across allied governments, and get red lines written into law and contract, not left to a policy a court fight has to defend.
• Build licensing-grade safety now, and sell it as market access. Governments are moving catastrophe-capable AI under mandatory licensing and screening — the lab heads themselves backed mandatory DNA-synthesis screening (ScreenDNA) — and sector regulators bind in 12–18 months (the FDA already at 1,451 AI devices, FDA). Enforcement is no longer theoretical: from 2 August 2026 the EU can issue information requests, demand access to a general-purpose model, and recall it from the market, with fines up to the greater of €15M or 3% of worldwide turnover, backed by 38 additional staff. The same date makes chatbot disclosure and synthetic-content marking enforceable across a 450M-person market (TechPolicy.Press). Ship evaluation, provenance, and audit ahead of the threshold, so the rules that stall rivals clear you to sell.
• Take the distribution that antitrust is prying open. The European Commission ordered Google to open 11 Android features to rival assistants and to share anonymized search data with competing chatbots, with changes due by July 2027 (TNW). It has since opened formal antitrust cases into Google’s use of publisher and YouTube content for AI, and into Meta’s terms for AI access to WhatsApp (European Commission). Each case is a door a regulator may hold open for you. Mandated access is the cheapest distribution a challenger will ever be offered. Build the integrations now, so you are ready on the day the interfaces open rather than a year later.
• Capitalize on the sovereignty scramble. Dependence, not just competitiveness, now drives buyers. After the US export cut-off, France pledged €100B+ and Mistral built its own data center near Paris (TechCrunch), while the EU and India mobilize sovereign compute and models (G5). Offer sovereign deployments — on-shore weights, national-cloud hosting, local control — as a distinct line to squeezed-middle states that will pay a premium for autonomy.

Lens W — the World. Lens W is the system view. Competition narrows the field to 5–7 frontier firms, pools rent at the compute-and-model layer, and now registers as a stability risk — the IMF and BIS flag AI concentration and circular vendor financing (W1, W2). The likely near-term path is a capital-cycle bust that concentrates the field, and rising antitrust and unbundling pressure aimed at the roll-ups. For an AI firm, this decides which players survive the correction, and on what terms. Derived recommendations:

• Survive the bust that concentrates the field. Circular financing (NVIDIA↔︎OpenAI↔︎Oracle) makes demand look firmer than it is, and Bain sees AI ~$800B short of the revenue needed by 2030 (Bain). A correction clears the over-extended and rewards the cash-disciplined (A3). The first tremor arrived in July 2026, when Alphabet raised capital spending guidance and its shares fell 7% the same day, taking the other hyperscalers with it. A capital spending raise stopped reading as confidence. Keep balance-sheet slack, and avoid multi-year commitments priced at today’s frontier rates, so a funding freeze prunes your rivals, not you.
• Assume structural separation is coming, and position on the right side of it. The one brake aimed at concentration itself is a “Glass-Steagall for AI” — barring one entity from owning the frontier, compute, information layer, and political money at once — and the roll-ups (SpaceX–xAI–Anysphere) are what draw it (Vanderbilt). As a challenger, back the portability and interoperability mandates that pry open incumbent lock-in. As an incumbent, avoid the acqui-hire route that invites a break-up.
• Pick your open-weight posture deliberately — it is the swing factor for the whole system. Open weights cap how much rent can be extracted and are the printing press of this era (W2, W5), and Chinese labs already ride it to ~61% of routed token volume by May 2026 (Data Gravity). Meta shows the cost of drifting: it defaulted into open release without defending the position, and its share of routed tokens fell below 1%. Decide on purpose whether open release serves your position — commoditizing a rival’s layer, winning developer share — rather than defaulting into a closed model whose rent antitrust will target.
• Do not assume today’s profitable layer stays profitable. Value capture moved from infrastructure to the model layer inside 2026, lifting lab inference margins from 38% to over 70% while chip and foundry pricing stayed restrained (SemiAnalysis). A strategy anchored to where the profit sat last year is a strategy anchored to a moving object. Re-test the assumption every planning cycle, and keep the option to operate one layer up or down.

↑ Back to Table of Contents

Coda

AI is now building itself. The advantage comes with what is genuinely scarce — electricity you have contracted, memory you have secured, the customer’s context you hold, and the trust that lets a regulator or a buyer say yes. The prize is large because everyone is realizing the value of AI and is buying more of it. This will drive market concentration. By 2028, the fastest few will own what the rest depend on.

Make AI build AI. Get it in everyone’s hands. Offer cheaply today, but not tomorrow.

↑ Back to Table of Contents

Stay tuned for more insights on AI from the perspective of the other lenses on
Robert Bruckmeier’s private paper channel.

Please cite this work as:

Bruckmeier, Robert. The AI Dividend and the Deadline — What Drives AI Tech Companies. Robert Bruckmeier’s private papers (August 2026). papers.robertbruckmeier.com/ai-dividend-and-deadline/ai-company-view.html

Or use the BibTeX citation:

@article{bruckmeier2026aidividend,
  title   = {The AI Dividend and the Deadline — What Drives AI Tech Companies},
  author  = {Bruckmeier, Robert},
  journal = {Robert Bruckmeier's private papers, papers.robertbruckmeier.com},
  year    = {2026},
  month   = {August},
  url     = {papers.robertbruckmeier.com/ai-dividend-and-deadline/ai-company-view.html}
}

© 2026 by Robert Bruckmeier