The Operating Manual for the AI Revolution
The biggest industrial shift in modern history is already underway, and if you cannot see the full stack, you cannot see where the world is heading.
How $60B of AI revenue generates $700B of infrastructure investment
72 Rubin servers + 256 LPU chips = 700M tokens/sec — 350× Hopper throughput. Purpose-built for inference.
Blackwell (now) → Vera Rubin (2026, HBM4) → Feynman (2028, TSMC 1.6nm, silicon photonics)
Anthropic closed a $65B Series H at a $965B post-money valuation on May 28, 2026, roughly tripling its $380B February mark and passing OpenAI for the first time. More importantly for this report, it also passed OpenAI on US enterprise penetration — which is a better leading indicator than either valuation or a leaked run-rate.
The application layer is the fastest-growing segment of AI revenue.
Historically, in every major computing wave — mainframes, PCs, mobile, cloud — the application layer is where most value ultimately accrues. Infrastructure enables, but applications capture.
Source: company disclosures, press reports, DVC analysis. “Pure application layer” = companies with own distribution, excluding model providers and generative media.
Time for each software category to reach ~$60B in combined application revenue
Source: Bessemer Cloud Index, Menlo Ventures, DVC analysisTop AI application companies by annualized recurring revenue
API-heavy = Most revenue from API/platform | End-user = Most revenue from consumers/enterprise seats
The application layer is becoming a behavior layer. Revenue ranks one thing; usage habits rank another. ChatGPT still anchors the category — 244M desktop conversations in March 2026, up 55% year over year — but the scarce asset is the default workflow and the context that accumulates inside it, not the model behind the surface.
Claude is the breakout: 22M desktop conversations in March 2026, up 1,858% versus October 2025. And the habit is spreading across surfaces, not consolidating into one default — AI assistant tools reached 36% of desktop users and 23% of mobile users in Q1 2026. Meanwhile, AI is reshaping the surfaces beneath it: Google’s AI Overviews appeared alongside 46% of paid search ads in consumer credit cards in Q4 2025, up from 21% in Q2 2025.
Source: Comscore Q1 2026 AI Intelligence ReportCapital is flooding every layer, but one battleground shapes the rest. To understand power in AI, you have to understand the fight over the models themselves.
THE MODEL MARKET BECAME A BARBELL. Premium multi-day agentic tiers priced up to $10/$50, capable output priced down to $1–$6, and a third axis — token efficiency — that no price-per-token chart captures. 10+ companies at the frontier simultaneously. The bottom is commoditizing; the top is getting more expensive and more strategic. The default ChatGPT tier has quietly moved up to GPT-5.5 Instant, so the consumer baseline now sits well above 2025 frontier on AIME and MMMU-Pro. Open-weight models look close to proprietary ones on public benchmarks; held-out government evals still show a gap. Cost efficiency, especially from PRC labs, is real. And the benchmark table is a weaker moat signal than it was: with Gemini 3.5 Flash and Claude Opus 4.8, distribution, runtime, and governed tool access now matter as much as raw capability — the contest is model + runtime + distribution, not a leaderboard alone.
We are retiring the “Frontier of Restraint” framing. Prices moved in both directions in a single quarter. Anthropic pushed the top end up to $10/$50 for multi-day agentic work while Sonnet 5 and GPT-5.6 Luna pulled capable output toward commodity pricing. The middle thinned, and vendors started competing on tokens per task rather than only price per token.
GPT-5.4 / o3
$25B ARR
GPT-5.5
Apr 24, 2026
Claude Opus 4.8
#1 Chatbot Arena
Mythos
Enterprise preview
Gemini 3.5 Flash
#2 Chatbot Arena
Grok 3
$1.25T merged
Nova 2 Pro
Top reasoning on Bedrock
Muse Spark
Closed · 3.6B DAU
In July 2026 the largest open-weight models in the world are all Chinese. This is not a price story — closed pricing already collapsed to $1–$6. It is a story about scale and ownability: governments, health systems and vertical platforms that need to own the model rather than rent it now have frontier-scale open weights to build on.
Open-weight progress on public benchmarks does not yet survive uncontaminated evaluation. Independent NIST/CAISI held-out testing puts DeepSeek V4 Pro at an IRT Elo near 800 versus around 1260 for current US frontier — a real, meaningful gap. We do not convert that Elo spread into a month figure: the conversion is not in the source. Cost efficiency tells the opposite story: V4 Pro was cheaper than GPT-5.4 mini on 5 of 7 benchmarks. The clean read is that public evals look close, held-out evals show a real gap, and PRC cost efficiency is a separate, real advantage. Source: NIST / CAISI
Llama 4 Maverick
Community License
V3.2 / R1
Open Weight · MIT
Qwen 3.5 397B
Apache 2.0
GLM-5 744B
#1 OS leaderboard
Kimi K2.5 1T
99.0 HumanEval
Mistral Large 2
EU Sovereign AI
Nemotron 3
Open Agentic
Gemma 4
Apache 2.0
GPT-oss 120B
Apache 2.0
Step-3.5-Flash
97.3 AIME
M2.5 230B
80.2 SWE-bench
July 2026 was the first window in which vendors competed on tokens per task rather than only price per token. The three figures below are each a vendor’s own comparison against its own predecessor, on different tasks, with no common benchmark. They must not be read as a curve. We have deliberately not drawn one.
THE FOUR SCALING LAWS OF AI — Jensen Huang, Lex Fridman Podcast #494 (Mar 2026)
Bigger models + more data + more compute = smarter AI. The original scaling law.
Synthetic data, RLHF, fine-tuning, distillation. “We are no longer limited by data — we are limited by compute.”
100x+ compute at inference for multi-step reasoning. “Inference is thinking, and thinking is hard.”
Agents spawn sub-agents, use tools, create data. “It’s like multiplying AI. We could spin off agents as fast as you want.”
“Intelligence is going to scale by one thing, and that’s compute.”
Source: Lex Fridman Podcast #494, NVIDIA GTC 2025–2026Billion-dollar bets on people and contrarian theses. Zero revenue, zero products.
Built ChatGPT, DALL-E, voice mode. Now building multimodal agentic AI.
One mission: safe superintelligence. No products, no distractions.
LLMs hit a wall. Building world models that learn from reality, not language.
AI as connective tissue for human collaboration.
Novel RL for superintelligence. Three months old.
Open frontier lab. Western answer to DeepSeek. No model shipped yet.
Opening the black box. AI interpretability.
Spatial intelligence. 3D world models from images. ~30 people.
Caveat: Silicon Valley has been here before. Massive pre-product rounds sometimes build category-defining companies — and sometimes they don’t. Thinking Machines Lab lost its CTO and cofounders back to OpenAI within six months of its $2B seed. H Company (ex-DeepMind, $220M seed) lost 3 of 5 cofounders to “operational differences.” SSI’s Daniel Gross left for Meta. xAI lost all 11 cofounders by March 2026. The talent that makes these bets valuable is also the talent most likely to leave. The bet is real. So is the risk.
Every modality now has its own model race. The frontier isn’t just LLMs anymore.
Benchmarks are dead. Meta admitted Llama 4 was tuned specifically to score well on benchmarks — prompting a credibility crisis across the leaderboard. Models are now optimized for benchmarks rather than tested by them. The industry needs new evaluation methods: real-world task completion, user preference studies, and domain-specific assessments.
Source: Meta Llama 4 controversy (TechCrunch, Apr 2025), Scale AI SEAL benchmark initiativeCost collapsed at the commodity tier while frontier list prices split upward. On current published list prices (August 2026), the cheap-and-capable lane runs at $1/$6 per 1M input/output tokens for GPT-5.6 Luna, against $10/$50 for Claude Fable 5 at the agentic frontier — a 10× spread inside one quarter. We no longer print a single headline decline percentage: the two series we previously carried used different model classes, baselines and windows, and neither could be sourced to a current index.
Take last year’s frontier model: a model that performs similarly on benchmarks is now 500–700× cheaper to run. This collapse in inference cost is what enables the application layer explosion above — and why agentic workflows (which require 10–100× more tokens) are suddenly economically viable.
Source: Epoch AI (Mar 2025), a16z price index, OpenAI/Anthropic/DeepSeek pricing pages, DVC analysisBase generation is commoditizing. Value migrates to orchestration, inference optimization, and proprietary data.
As raw intelligence gets cheaper, the center of gravity shifts. Value moves from generating answers to getting work done.
This is a structural change, not a number update. SpaceX signed an agreement to acquire Anysphere (Cursor) for $60B all-stock on June 16, 2026, subject to closing conditions. The coding application and frontier-model ambitions now sit inside one public-company structure. That collapses the clean separation between the model layer and the application layer that this report has used since its first edition.
Lovable reached $500M ARR in June 2026 (from $400M in February) on 146 employees — roughly $3.4M of ARR per employee, which is the cleanest quantitative evidence in this report that AI-native companies operate at a different cost structure. 8M users, 50M+ projects, ~1M new projects per week, ~$20M enterprise ARR (Uber, HubSpot, Microsoft). A $300M raise at a reported $13.2B valuation led by Menlo Ventures is reportedly in progress — roughly 26× ARR.
In talks to raise at $5B valuation — not closed. Separately reported at >$500M annualized revenue run-rate and cash-flow positive, which is unusual at this stage of a generative-video business and is the reason the round is being discussed at that level.
Sources: The Information (in-talks valuation) · Business Insider (run-rate, cash-flow positive)Voice now has the same shape the coding category had a year ago: a scaled independent, a frontier-platform entrant, and enterprise penetration you can count. It earns a section.
AI is moving from a system you consult to a system that acts. That changes software from a tool for humans to a layer of labor that can execute across workflows. The agent ecosystem alone has already created 67,000+ engineering openings globally — more than at any point in three years.
>$450M ARR
Answer engine → agentic platform
Transforms a spare Mac mini into an always-on AI agent that controls apps, browses the web, manages files. 19-model orchestration. Personal Computer product launched Feb 2026 — the first consumer device-as-agent play from a search company.
Source: Perplexity / The Verge, 2026$2.5B+ run-rate
9-month ramp to billion-dollar product
GA May 2025 → >$2.5B run-rate by Feb 2026 (most recent disclosed). Weekly active users doubled since Jan 2026. Business subscriptions quadrupled. Terminal-first agentic coding drove Anthropic to a $380B post-money Series G ($30B raised, Feb 2026) — since superseded by the $65B Series H at $965B post-money on May 28, 2026.
Source: Anthropic Series G announcement, 2026~$125M run-rate, breakout 2025
General-purpose agent · Singapore
Founded in China, moved to Singapore. Reported $100M+ ARR in 8 months, ~$125M run-rate by late 2025, 147T tokens processed, 80M+ virtual computers, with Windows 11 trials at Microsoft. The general-purpose agent category is becoming a strategic acquisition target for hyperscalers; a Meta acquisition has been discussed in earlier reporting but we are not citing it as confirmed here without a primary source.
Source: CNBC / WSJ / AP News (Dec 2025) · treat any Meta acquisition framing as unconfirmed pending a primary release$4B ARR (Jun 2026) · $60B SpaceX deal signed Jun 16 2026, expected to close Q3 2026
$1B run-rate (late 2025) → $2B+ (Feb 2026)
Fastest SaaS growth curve in history. ~60% revenue now from enterprise (was individual-first). 50,000+ enterprises, 100M+ lines of enterprise code per day. $500M ARR Jun 2025 → $1B late 2025 → $2B Feb 2026. Apr 21, 2026: SpaceX (xAI) announced $10B partnership investment with option to acquire Cursor outright — exercised Jun 16, 2026, and the transaction is signed, not closed, with completion expected in Q3 2026 — for $60B — Cursor gains Colossus supercomputer access, resolving its “bottlenecked by compute” constraint.
Source: Bloomberg / TechCrunch, 2026 · SpaceX announcement Apr 21, 2026$10.2B valuation
67% PR merge rate — autonomous SWE
First fully autonomous software engineer. Devin 2.0 handles async multi-step tasks: reads codebase, plans approach, writes code, runs tests, submits PRs. Moving from code completion paradigm to autonomous project execution.
Source: Cognition, 2026Jensen-era hype → absorbed by Anthropic
Open-source Claude wrapper, peak Mar 2026
Rise (Mar 2026): Jensen Huang at GTC 2026 called it "as big as HTML, as big as Linux" — an open-source personal AI agent runnable on Mac mini ($599), RTX PCs, DGX Spark, DGX Station, or cloud VPS (~$30/mo). Star count at peak is not printed: the figure we previously carried could not be sourced to GitHub.
Fall (Apr 2026): Systematically absorbed by Anthropic over about four weeks — trademark warning, OAuth blocked, key features cloned, then Anthropic's "Channels" subsumed the last differentiator. The creator left to join OpenAI. Treat OpenClaw as a past-tense reference unless a successor fork takes its place.
Source: NVIDIA GTC 2026 keynote / Business Insider / TechRadar (Mar 2026) · The Register / Hacker News / GitHub (Mar–Apr 2026)Enterprise AI Agent Layer
NVIDIA's enterprise wrapper for OpenClaw
NemoClaw (NVIDIA, GTC 2026): Enterprise security layer — network guardrails, privacy router, sandboxed execution via OpenShell. Installs with a single command. Adds Nemotron models + Dynamo inference engine. Jensen's pitch: "OpenClaw for everyone, NemoClaw for the enterprise."
Source: NVIDIA GTC 2026 keynote / TechRadar, Mar 2026Jensen Huang declared the arrival of the "inference inflection point" at GTC 2026: two exponentials colliding — demand for inference growing exponentially while cost per token falls exponentially. The question is no longer whether agents can work. It's whether the business models can sustain them.
We unpack the full business model problem — pricing paradigms, margin squeeze, and why the economics are still unresolved — later in the presentation.
Source: NVIDIA GTC 2026 keynote / Axios / CNBC, Mar 2026"Ability to make software will be a human right soon, and it's not going to feel like making software."— The Vibe Coding Thesis
AppDirect: Non-technical marketing team vibe-coded 200K+ lines of code, built 11 projects with 4 in production, and have 80+ applications in progress across Sales, Finance, HR, and Operations.
Zero-code founder: Built a transcription platform that reached 80,000 users, 1M+ minutes processed, and six-figure ARR — in four months.
Source: Lovable / AppDirect case study, Replit / Whisper AI, 2025Agents building agents: both Anthropic and Perplexity say their coding tools are now built with themselves — as is this presentation. The “100%” figures we previously printed are vendor claims we could not verify and are not literally checkable.
The nature of code itself is changing. Humans write abstractions — functions, classes, design patterns — primarily so other humans can read and maintain the code. AI does not need that. It can generate and re-generate from scratch faster than it can navigate a complex abstraction hierarchy. Code is becoming a throwaway artifact rather than a maintained asset. 42% of all committed code is now AI-generated or significantly AI-assisted (SonarSource, survey of ~1,150 developers, Jan 2026); developers themselves forecast 65% by 2027. The same survey’s headline finding is a verification gap: 96% do not fully trust that AI-generated code is functionally correct, and only 48% always check it before committing.
Source: SonarSource State of Code 2026But the ceiling is rising faster than the floor. While vibe coding democratizes building, advanced practitioners are diverging fast. Anthropic’s 2026 Agentic Coding report: “Software development is shifting from writing code to orchestrating agents that write code.” Engineers now run multiple AI agents in parallel on one codebase (Vibe Kanban, AutoForge), each on isolated git worktrees, with visual kanban boards for task management. Claude Code’s dynamic workflows push this further — one prompt fans out into hundreds of parallel subagents. The new SWE job: decompose tasks, spin up agents, review their PRs, resolve merge conflicts. A tech lead managing a team of AI juniors. One company deployed 800+ AI agents internally.
The throughput signal is real before the headcount signal is: code-contribution volume has risen far faster than US software-developer employment. We no longer print the two specific percentages we previously carried here — neither could be sourced to Microsoft Research or to the BLS. AI is changing throughput and the unit of work faster than it is clearly compressing headcount — the work is shifting shape before the labor market does.
Sources: Anthropic 2026 Agentic Coding Trends Report · Microsoft Research — Diffusion of AI in Software DevelopmentCompute is the new acquisition currency. SpaceX (which absorbed xAI) has put a deal on the table with Cursor: either invest $10B in the partnership or acquire Cursor outright for $60B. Cursor gets access to Colossus, xAI’s supercomputer. Cursor’s CEO explicitly said they were “bottlenecked by compute.” The model layer is starting to buy the application layer, and the weapon is not cash, it is GPU access. The first time an infrastructure player has used compute itself as acquisition currency for an app-layer company. Expect more of these.
Source: SpaceXThe agent stack is no longer startup middleware — it is becoming governed execution infrastructure. Hyperscalers and enterprise incumbents are turning agents into managed MCP servers, identity-scoped tools, audit trails, sandboxed code execution, shared operational context, and systems of action. The frontier has shifted from “can an agent answer?” to “can an agent safely do work inside the enterprise?” The product surfaces above the seven-layer base now look like this:
The A2A protocol has crossed 150+ organizations with native integration across Google, Microsoft, and AWS, complementing MCP as the cross-org agent-to-agent layer. The pattern: above the 7-layer agent stack, a governance and control-plane layer is forming, and the hyperscalers and enterprise incumbents are racing to own it.
Source: Microsoft · Cisco · AWS · ServiceNow · Google Cloud · AnthropicHow fast does open source move? In late March 2026, Anthropic accidentally shipped Claude Code’s source in an npm package. The community archived it and reimplemented the core in other languages within days. We no longer print the line count, fork count or the names of the unreleased features the leak was said to expose: none of those four specifics could be verified.
Meanwhile, OpenClaw showed how quickly an open-source wrapper can force a platform response: trademark and authentication friction arrived, overlapping features moved into Anthropic’s own product, and the creator later joined OpenAI. We no longer print volatile GitHub-star counts or a four-week causal sequence that could not be independently reconstructed.
The moral: proprietary code is a temporary state. Once a useful interface reaches the open internet, the community can absorb and reimplement it rapidly. Open source does not merely compete with proprietary software; it shortens the half-life of differentiation.
Source: The Register, Hacker News, GitHub, Mar 31, 2026ChatGPT Agent, Devin, Manus
Subscription / usage-based
Internal infra, VPS, hybrid
Higher ops, better governance
Cursor, Claude Code, Copilot
Software subscription only
Perplexity Mac mini, OpenClaw, DGX Spark
$599+ one-time + low variable
FASTEST GROWINGCloud held 81.1% of agent market share in 2025 — but local-first is the next visible deployment wave
Source: Mordor Intelligence, 2025The next leg of the agent build-out is not a better chat UI. It is the Palantir FDE playbook at frontier-lab scale, on two fronts:
Frontier labs are moving beyond model APIs into implementation services and partner vehicles. Exact fundraising, valuation and portfolio-client figures previously printed here were removed after the linked report became unavailable and could not be independently reproduced.
Source: TechCrunchAnthropic now ships 10 ready-to-run finance agents (pitchbooks, KYC, month-end close, and more) via Claude Cowork, Claude Code plugins, and Managed Agents, with Excel, PowerPoint, Word, and Outlook add-ins inbound and a Moody’s MCP app covering 600M+ public and private companies. Generic chat is giving way to packaged vertical workflows.
Source: AnthropicYou must make yourself vulnerable to extract value — but that will change.
Snyk ToxicSkills: 37% of OpenClaw community skills contained flawed code. 200+ GitHub security advisories.
EchoLeak attack on browser agents. Slack AI data exfiltration. Gemini memory poisoning demonstrated.
Agents need files, email, calendar, purchases to be useful. Every permission granted is an attack surface.
Today: accept risk to capture value. Tomorrow: agent-specific security layers, capability-based permissions, cryptographic identity.
Agents now see your screen. Claude Computer Use lets Claude see your desktop, launch apps, browse the web, and fill spreadsheets. OpenAI introduced native computer use in the GPT-5 generation and continues it in GPT-5.6. OpenClaw brought the pattern to open source. Agents no longer need a custom API for every tool — they operate software through the same interface you do. That is a massive unlock for automating legacy systems that will never get an API.
Source: Anthropic, OpenAI, Mar 2026Agent scans kids' school emails. Finds early dismissal — short day today.
Sends iMessage to nanny: "Short day — pickup at 12:30 instead of 3."
Cleaning lady texts via Telegram: "You're out of garbage bags."
Agent orders from Amazon using its own account and crypto card. No human needed.
Checks weather forecast, pre-cools house for afternoon heat via HVAC.
Adjusts lighting, unlocks door via HomeAssistant, confirms to parent.
Reviews tomorrow's calendar, preps grocery list, charges batteries at off-peak rates.
It’s not just tech companies. A roofing company is using AI agents to pull satellite imagery, cross-reference hail damage patterns, and feed warm leads to their sales team. They’re roofers — not engineers, not a startup. When a roofing company runs AI agents, every company runs AI agents.
Source: @RoundtableSpace / Startup Ideas Podcast, Apr 2026CHEAPER AI ≠ LESS SPEND. Cheaper inference = more workflows clear the ROI threshold.
Task completion sounds simple until you look under the hood. What appears to be one product is really a new software stack in disguise.
Every agentic product — from Cursor to Harvey to Glean — is built on the same fundamental layers.
Click any layer to explore tools, companies, and key data. Hover any company for details.
Agent ↔ Tools / Data
"The REST of AI" — how agents access external tools and data.
Anthropic-led. Adopted by OpenAI, Google, Microsoft.
Agent ↔ Agent
How agents collaborate across organizations.
Google-led. Salesforce, SAP, ServiceNow.
Agent ↔ User
How agents surface work to humans. CopilotKit-led open protocol.
Supported by LangGraph, CrewAI, Microsoft, Google, AWS.
"Together, these three protocols are creating an interoperable agent ecosystem — the TCP/IP moment for AI agents."
The enduring advantage in agents will not come from having a model. It will come from orchestrating the full system around the model: memory, tools, workflows, reliability, and distribution. We have moved from prompt engineering to context engineering.
That architecture is the map of defensibility. The winners will not be the loudest at the frontier; they will be the ones who turn intelligence into dependable, repeatable execution.
If today's economics are strained, the next interface may rewrite them. The moment software starts transacting with software, the market changes shape again.
We've built agents that talk to humans. The next frontier: agents that discover, hire, pay, and supervise each other.
Moltbook — a Reddit-like social network for AI agents — went from niche experiment (Jan 2026) to Meta acquisition (March 10, 2026) in ~6 weeks. Agents posted, commented, upvoted, and gossipped about their human owners.
Probably not. The durable market looks less like "bots posting on bot Reddit" and more like a programmable service economy — authenticated agents discovering each other, negotiating work, moving money, and leaving auditable trails.
Agents can't open bank accounts, pass MFA, or handle card fraud flows. But they can hold programmable balances and transact instantly.
The "body" isn't a humanoid robot — it's a dedicated machine running 24/7 with local files, apps, and persistent memory.
The Mac mini is emerging as the default "agent hardware" — cheap, quiet, always-on, with enough local compute to be a persistent digital worker.
The agent-to-agent economy is real enough to invest in, but early enough that the biggest winners may not be the agents themselves — they may be the companies that provide the protocols, identity, payment rails, and trusted execution environments that let agents safely discover, hire, pay, and supervise one another.
That software stack still runs on steel, silicon, and electricity. The more capable AI becomes, the more brutally physical the system underneath it gets.
Every breakthrough at the application layer is paid for in chips, data centers, cooling, and power. AI is driving the largest infrastructure buildout since the interstate highway system, with capital racing ahead of certainty. The bottleneck is no longer just compute; it is whether the physical world can support the pace.
Top-four 2026 capex guidance now sits at ~$725B at the midpoint, up to ~$745B at the top end, against roughly $402–410B in 2025 (+77%): Amazon ~$200B, Microsoft ~$190B, Google raised to $195–205B, Meta $125–145B. Including leases, calendar-2026 capex approaches ~$800B. That part of our May thesis survived. What changed is that hyperscalers stopped funding the buildout out of cash flow — which converts an operating-leverage story into a balance-sheet-leverage story and materially raises the whole stack’s sensitivity to a demand air pocket.
Our May edition assumed an effectively unchallenged NVIDIA. Three independent breaks landed in the same quarter.
“Microsoft restarted Three Mile Island” is the most-repeated power datapoint in AI and it is misleading as usually told. The Crane Clean Energy Center restart (835 MW, 20-year PPA, ~$16B) produces no power until H2 2027 at the earliest — commercial operation is now targeted for H2 2027, pulled forward from 2028. We no longer print a FERC waiver date: it could not be sourced. Sector-wide the gap is structural, and the bridge is gas.
Other in-window power moves: Google–NextEra agreed to restart Iowa’s 615 MW Duane Arnold plant (a ~$1.6B restart, as described earlier in this report). Federal loan-guarantee activity and a widely reported Kentucky campus were also in the window, but the specific figures we previously printed could not be sourced and have been removed rather than restated. Jensen Huang’s own framing is that the binding constraints are now HBM supply, land, electricity and construction labour — not GPU fabrication.
Sources: SMR Intel tracker (May 2026 cut) · DistroForge · AI Power Weekly (Project Kilby) · Broadband Breakfast (DOE) · Data Center Dynamics · Yahoo Finance (NextEra)Latest reset (through July 2026): hyperscaler 2026 capex guidance has stepped up to a ~$725B midpoint / up to ~$745B top end — about $100B above the projections we built this section on. Per-company anchors used in the CapEx explorer below, and kept consistent across the report and the deck: Amazon ~$200B, Alphabet ~$200B ($195–205B), Microsoft ~$190B, Meta ~$135B (midpoint of $125–145B), Oracle ~$50B. The infrastructure chapter is still getting bigger; the bottleneck is memory, power, and data-center execution, not appetite.
Source: Yahoo Finance / Business Insider, Apr 29–30 2026Capacity is not abstract capex — it sets the product. Anthropic signed for all compute at SpaceX’s Colossus 1 — 300+ MW and 220,000+ NVIDIA GPUs — and immediately raised Claude usage limits. Rival labs renting each other’s data centers is the clearest proof that compute scarcity, not model design, is what currently caps rate limits and product experience.
Source: Anthropic2026 CapEx will consume ~94% of operating cash flows — vs a 10-year average of 40%. For the first time, hyperscalers collectively hold more debt than cash.
Purpose-built AI infrastructure at hyperscale
Source: Nebius, 2026GPU cloud built for AI-first workloads
Source: CoreWeave S-1, 2025Jensen's core reframe: data centers aren't storage facilities anymore. "Electrons go in, tokens come out." The $700B CapEx sprint is building the world's first generation of AI factories — purpose-built for inference at scale.
Anthropic has contracted for the full compute capacity of SpaceX’s Colossus 1 in Memphis (300+ MW), with reported interest extending into space-based compute. The AI factory has become fungible enough that competitors rent it from each other.
Source: CNBCThe NVIDIA / IREN deal is now the template: a $3.4B five-year managed GPU cloud contract at Childress, Texas; a five-year warrant for up to 30M IREN shares at $70 (about $2.1B notional, struck above the announcement-day close); and an intent to build up to 5 GW of DSX-aligned AI factory capacity, with the 2 GW Sweetwater campus as the flagship (Sweetwater 1 at 1.4 GW already on ERCOT). DSX is the reference architecture for AI factory construction.
Source: NVIDIAMicrosoft remains OpenAI’s primary cloud partner, but the relationship is non-exclusive: OpenAI can serve products on any cloud, the Microsoft IP license through 2032 is non-exclusive, OpenAI’s revenue share to Microsoft continues through 2030 (capped), and Microsoft no longer pays revenue share to OpenAI. Distribution and compute are multi-cloud by default.
Source: Microsoft / OpenAIFY2026 Revenue — +65% YoY
$1T+ in Blackwell + Vera Rubin orders through 2027
Source: NVIDIA GTC 2026 keynote / CNBC / Axios, Mar 2026Groq LPU → $20B non-exclusive technology licence plus asset purchase and acqui-hire (Dec 2025); NVIDIA states it did not acquire Groq. NVIDIA now folds the licensed inference technology into its own architecture. Senators Warren and Blumenthal are probing whether the structure evaded merger review.
China generates substantially more electricity than the US in total and has been adding capacity at a rate the US has not matched, and is projected to carry large spare capacity into 2030. We no longer print the specific capacity-addition or spare-capacity figures: neither could be sourced to the IEA or to a named analyst publication. Energy experts who visit China describe power availability as a "solved problem."
Yet the US consumes nearly 2× more data center electricity. The asymmetry: the US has the chips but is hitting energy bottlenecks (sell-side forecasts point to a multi-tens-of-gigawatts shortfall by 2028; we no longer print the specific figure, which could not be sourced). China has the electrons but is constrained by US export controls on high-end GPUs.
The race for AI supremacy may not be won by who builds the best model — but by who solves their bottleneck first.
Source: Brookings (Feb 2026), Fortune, IEA, Morgan Stanley$500B commitment
Joint venture with SoftBank, Oracle, OpenAI
Delays and scope adjustments
Power procurement bottlenecks
1.2GW Abilene campus
First phase operational
As AI shifts from training to inference (Jensen's "inference inflection point"), power demand doesn't decrease — it redistributes. Training clusters run in bursts; inference runs 24/7. Always-on inference = always-on power demand.
Source: NVIDIA GTC 2026 / industry analysisDispatchable MW with an executable timeline is the scarce input.
Once autonomous systems can coordinate digitally, the next step is obvious. They move out of chat windows and into the real world.
Any framing that implies imminent convergence is now wrong. Waymo scaled; Tesla’s paid mileage was flat across two quarters while adding cities.
Humanoid robotics funding reached $8.6B in 2026 by late July — 1.8× the entire 2025 total, with five months left in the year (Dealroom). That gives this section comparables it did not have in May. It also inherits the sector’s unresolved problem.
>$450M raised · >$2.4B valuation. Two-arm general-purpose robot, using video models for physical imagination rather than hand-coded motion policies. Figures per DVC.
"The moment when a robot can do everything better than a human doesn't come once in a decade or once in a lifetime, it happens once in the history of humanity, and we're close to it"— Andrew Wooten, CPO, Rhoda AI
Foundation models powering physical AI
The same transformer architecture powering ChatGPT is now learning to control physical robots
Autonomous vehicles are no longer a concept — they're on the road
$16B raised (largest AV round ever). 11 public driverless metros plus four more in employee-only operation as of Jul 8 2026. 220M+ rider-only miles through end-March 2026. The 1M rides/week year-end target needs a 2× from ~500K, which has been flat since March.
Austin launch June 2025. FSD: 1 collision per 5.3M miles vs national avg 1 per 660K. Cybercab production 2026, <$30K by 2027. Fully driverless tests began Dec 2025.
First commercial driverless permits (Aug 2022). Operates in Wuhan, Chongqing, Shenzhen. Expanding to Abu Dhabi.
Purpose-built robotaxi with no steering wheel. Testing in SF, Vegas, Foster City. Las Vegas as first commercial market.
IPO'd. Operates in Shenzhen, Shanghai, Beijing, Guangzhou.
IPO'd. ~150 cars across Abu Dhabi, Dubai, Riyadh. Middle East expansion.
$750M Series C (Jan 2026). Strategic Uber partnership; the previously printed robotaxi-volume target was removed because it could not be independently sourced.
$1.2B Series D (Feb 2026). SoftBank and NVIDIA backed. End-to-end learned driving.
Nebius subsidiary (ex-Yandex SDC). Live robotaxi on Uber in Dallas. Delivery robots on Uber Eats in 3 cities. Building both AV and last-mile delivery.
8× safer than human drivers — 1 collision per 5.3M miles Source: Tesla Safety Report, 2026
Self-driving trucks are commercially hauling freight on US highways. The $1T US trucking industry is the first autonomous market generating real contracted revenue.
1,000-mile Fort Worth→Phoenix route. Partners: Volvo, PACCAR, FedEx, Uber Freight, Werner. Targeting ~$1B rev by 2030.
First US company with fully driverless trucks at commercial scale (Jan 2026). Fortune 50 retail customers. Partners: Isuzu, NVIDIA, Ryder.
Interstate runs from Texas hub. Customers: J.B. Hunt, Werner Enterprises. Also developing autonomous defense vehicles for US military.
Global labor market by sector — and what's automatable
The infrastructure buildout is breathtaking. But even with $700B flowing in, the fundamental economics of AI are still being figured out.
In May this was a directional thesis. By July a major incumbent had published outcome pricing and bought the metering infrastructure to bill it; Gartner had quantified the spend at risk; and a global services firm reported contracts moving to outcomes. This section got stronger, but the audit removed vendor prices and survey statistics that were not traceable to primary sources.
Bret Taylor’s framing is now the category’s thesis statement: “Outcome-based pricing is the future of software business models. The atomic unit of AI productivity is a process, not a person.”
AI is already reshaping how industries operate, compete, and ship product. Yet many of the companies building the core technology are still burning cash to deliver that transformation. The demand is undeniable; the economics are still unresolved.
Subscription caps upside and mis-prices heavy users. Usage / per-token / per-work-unit pricing captures expanding task volume and compute intensity. Anthropic 3× in 4 months on per-token API. Perplexity +50% in one month after Computer launched (per-task work). OpenAI carries 900M+ WAU but only 5.6% pay — subs are distribution, not value capture. Compare blended ARPU:
Subscription caps upside; usage scales with workload. Anthropic ARPU $16.20/mo (vs OpenAI $2.20, Google $1.10), premium per-token customers, not free-tier reach. Per-token revenue grows with task volume and compute intensity; flat subs mis-price heavy users. Coatue’s services-as-software framing projects a TAM expansion from a ~$0.2T software market into a ~$5.5T services-as-software paradigm, but only 3.8% of traditional SaaS spend is consumption-based so far. Most production agent companies are on hybrid pricing (a subscription floor plus usage and outcome components), not pure outcome billing. Outcome pricing is the strategic direction, hybrid usage is the current commercial reality, and subscriptions function as a revenue floor inside hybrid models rather than the value-capture engine.
Source: Counterpoint Research, Anthropic disclosures, Perplexity, OpenAI, Statista (Meta ARPU 2025), Netflix Q4 2025 earnings, Coatue C:\Takes, Orb (2026 State of AI Agent Pricing)The next trillion-dollar company might run on a business model we haven't seen yet. Subscriptions, usage-based pricing, ads, marketplaces, agent-to-agent payments — AI is still auditioning revenue models, and the winners may not look like any software company that came before.
Pay per token consumed. Scales with workload — today's economic engine.
Charge for resolved tasks, not raw compute. Aligns cost with value.
Useful on-ramp / distribution bundle — not the value-capture engine.
Pricing is migrating toward the unit of work — and now toward the unit of action. ServiceNow meters headless agent actions in the same Assist currency it uses for human work; AWS charges for the resources an agent consumes rather than for the toolkit itself. The market is experimenting with subscriptions, usage, and executed-work pricing — no universal model has settled.
Reality check: OpenAI’s $100M sounds impressive — until you do the math. That is ~$0.12 per user per year. Google makes ~$60. Fewer than 20% of eligible users see ads daily. AI ads may fund free tiers, but they cannot fund inference at scale.
Eyeballs vs wallets. Counterpoint’s Q1 2026 model estimates materially higher revenue per user for Anthropic than OpenAI, Microsoft or Google. The precise ARPU values are not repeated here because the underlying user denominators mix MAU and WAU estimates and conflict with Sensor Tower’s Claude MAU estimate. The directional point remains: premium enterprise usage monetizes differently from consumer reach.
Source: Counterpoint Research / The Register, Apr 30 2026A new pattern: branding restraint. Anthropic restricted Mythos to enterprise and US government cyber defenders. Days later, OpenAI restricted GPT-5.5-Cyber to the same audience with the same framing: “too dangerous for public release.” Both labs are now competing on what they don’t ship. National security access becomes the moat. Withholding capability becomes the brand. Every frontier lab from here will run the same play.
Source: The Verge, NYTOpenAI killed Sora after a short consumer run. The associated studio-licensing experiment ended with it. IP and licensing remained difficult enough that even the largest AI company continued pruning products rather than treating every launch as permanent.
Agentic AI as a Service — Jensen's framing at GTC 2026. Agents don't just answer questions — they complete workflows. Pricing shifts from per-token to per-task. Every SaaS vendor becomes an AaaS vendor — or gets disintermediated by an agent.
Pay per seat → Pay per token → Pay per outcome
Pay per workflow completed. The agent IS the product.
Where AI autopilots are attacking services. Tap any category for AI contenders, funding dates and sources.
This is where the agent control plane meets the P&L. Governed action layers — identity-scoped tools, approvals, audit trails, and liability boundaries — are what make agents legible to enterprises, and legibility is what lets them attack services budgets, not just software budgets. ServiceNow’s Action Fabric (governed system of action) and Cisco’s Cloud Control (humans and agents operating infrastructure from one data layer) are the canonical examples: AI moving into enterprise work execution, not just SaaS copilots.
The agent-to-agent economy is real enough to invest in, but early enough that the biggest winners may not be the agents themselves — they may be the companies that provide the protocols, identity, payment rails, and trusted execution environments that let agents safely discover, hire, pay, and supervise one another.
Pricing noise aside, some industries reward AI more than others. Healthcare is the largest, the most regulated, and the most structurally strange — and it is the one place where we can now measure what happens when AI meets a payment model that rewards documented intensity.
Healthcare AI is not one market. It enters a $5.3T system where the buyer, user, payer, and beneficiary are often different entities. It is a set of regulated loops with shared data and misaligned incentives — adoption is already arriving from patients and clinicians, while liability, reimbursement, and state-level rules decide where value can actually be captured.
Follow the money to see where value can be captured. Follow the patient event to see why adoption is shaped by workflow, reimbursement, data, and trust.
This is the thesis of the section, and between May and August 2026 it stopped being an argument and became a measurement. Both arrows below are true at the same time. That is the whole point.
Our May edition argued that reimbursement gates adoption. That was half the answer. Reimbursement decides what gets bought. Liability decides what gets deployed. And a single “10–15 year transformation horizon” is unfalsifiable — it was contradicted in both directions inside this window.
Modeled allocation constrained to official 2024 CMS NHE node totals. CMS does not publish a complete payer-to-service or service-to-cost-pool matrix.
| Channel | Value |
|---|---|
| Private health insurance | $1,644.6 |
| Medicare | $1,118.0 |
| Medicaid | $931.7 |
| Out-of-pocket | $556.6 |
| Other third-party payers & programs | $590.5 |
| Other NHE / reconciliation | $458.6 |
| Destination | Value |
|---|---|
| Hospital care | $1,634.7 |
| Physician & clinical services | $1,109.7 |
| Retail prescription drugs | $467.0 |
| Other health, residential & personal care | $320.5 |
| Nursing care facilities & CCRCs | $219.9 |
| Dental services | $189.2 |
| Other professional services | $184.9 |
| Home health care | $169.4 |
| Other non-durable medical products | $128.7 |
| Durable medical equipment | $86.4 |
| Admin, public health, investment & other | $789.6 |
Fewer, bigger, later — the same regime we describe at the frontier, one layer down.
Node totals match official CMS 2024 NHE data. Internal routing and operating-cost decomposition are modeled allocations constrained to those official totals.
More than eight companies are now operating at the frontier. The question is no longer who can build — but who owns the full stack.
Everyone says AI is built on three pillars: Data, Compute, and Talent. They are right — but incomplete. Talent remains the scarcest resource: AI roles are exploding, with a third of all openings concentrated in the Bay Area. Yet the conventional model describes the machinery. It does not describe what makes the machinery work.
Without distribution, the best model is a science project. Whoever controls the surface — search, devices, social, enterprise — chooses which models users touch.
Proprietary loops = moats
Inference overtaking training
74% of startups report inference-dominant costs
$10–20M/yr for top researchers
Best model + slow shipping = loss. Culture is the conversion rate of every other input. Speed of execution is the secret ingredient.
Every giant has unmatched resources — and a critical vulnerability.
Startups win in the seams.
Google processes ~15B queries/day. Perplexity hit 200M daily queries by mid-2025 (~1.3% of Google), up from 30M at the start of the year — targeting 1B/week. ChatGPT handles billions of prompts a day across a user base OpenAI last disclosed at 900M weekly actives (Feb 2026). Combined AI search share is growing 20%+/month. The sharper risk: AI answer engines reduce high-intent query volume and compress ad inventory economics even if overall share holds.
Google's proprietary TPU stack saves an estimated ~$3B/year vs. third-party compute for AI-augmented search. Ironwood (7th gen TPU) is "the first designed specifically for inference at scale" with 10× compute improvement and 2× power efficiency vs. prior high-perf TPU.
Gemini app: 950M MAU (Jul 2026, up from 750M in Feb 2026). Google Cloud: 48% growth in Q4 2025, run rate above $70B. All 15 products with 500M+ users now use Gemini models. 8M+ paid Gemini Enterprise seats sold in 4 months. Reuters called Google the AI momentum leader in Feb 2026.
Google invested $3B+ in Anthropic and committed up to $40B more (Apr 2026). Anthropic trains and runs Claude on Google TPUs — just expanded to multi-GW deal with Google + Broadcom for 3.5 GW of next-gen TPU capacity starting 2027. Google earns strategic exposure and infrastructure revenue from a frontier lab at a $47B run-rate (reported Jun 2026).
$201B revenue in FY2025, $60.5B net income. 3.58B Family Daily Active People. Meta can stuff AI into Facebook, Instagram, WhatsApp, Messenger, ad tools, smart glasses and creator tools across billions of users — even with a weaker model.
Llama 4 received poor reception (benchmark-tuning controversy). DeepSeek seized the open-weight lead. Meta planned its fourth AI restructuring in six months by Aug 2025. Cut ~15,000 jobs (20% of workforce) while simultaneously projecting $125–145B in AI CapEx for 2026 (raised from $115–135B in late April; stock dropped 7% after-hours on the news) — the clearest signal yet of replacing headcount with compute. Response: $14.3B investment in Scale AI (49% stake, no voting rights). Scale AI founder Alexandr Wang joined Meta to lead the Superintelligence Lab. Separately, Meta recruited Andrew Tulloch (co-founder of Thinking Machines Lab with Mira Murati) — reportedly offering up to $1.5B in compensation over 6 years.
Meta hired Nat Friedman and Daniel Gross to lead the reorg. Brought in talent from DeepMind, OpenAI, Anthropic. Sam Altman said Meta offered OpenAI employees $100M bonuses. The comp war is real but culture stability remains the question.
100M+ MAU across Copilot apps. M365 Copilot drives real ARPU growth across Office, Azure, and GitHub. Multi-model platform (OpenAI, Anthropic, Mistral) gives enterprises flexibility no other cloud offers.
Only 2.4M daily Copilot web visits vs. ChatGPT’s 400M+. Bing AI never broke through. Consumer identity is invisible. The multi-model strategy is a platform strength but a brand weakness.
OpenAI dependency is real: if OpenAI builds its own cloud, Azure loses its biggest AI differentiator. Microsoft hedged by signing Anthropic and Mistral, but OpenAI is ~70% of AI workloads on Azure.
Best on-device inference silicon in the world (M4, A18 Pro). 2B+ active devices. On-device AI processing could be the privacy moat that no cloud company can match.
Apple Intelligence received “mostly underwhelming” reviews. Siri overhaul delayed a full year. Tim Cook “lost confidence” in AI leadership. Switched from OpenAI to Google Gemini for redesigned Siri stack (Jan 2026).
SpaceX acquired xAI in Feb 2026 in the largest merger in history. Combined valuation: $1.25T. SpaceX valued at $1T, xAI at $250B. The IPO completed on June 12, 2026, reported at $135/share and ~$75B raised, so the $1.25T combined private mark now sits behind a public market cap.
X provides 500M tweets/day as real-time data flywheel. Grok 3: 93.3% AIME 2025. SpaceX generates ~$8B profit (50% margin). Filed plans for orbital AI data centers with up to 1M satellites.
No enterprise AI playbook. Consumer Grok traction unclear vs. ChatGPT/Claude. Orbital data centers are 2–3 years out. But: vertical integration of AI + rockets + satellites + data is unprecedented.
~$1B invested across 50 AI deals in 2024. By 2025: up to $100B committed to OpenAI, $10B to Anthropic, $2B to xAI. Total disclosed ecosystem financing: $33.8B in rounds NVIDIA participated in.
Back demand creators (model labs, apps) while also backing supply-side lock-in (infrastructure, networking, robotics). Foundation models: $20.9B in round sizes. Apps: $5.3B. Cloud/infra: $3.5B. Robotics: $3.3B. Even fusion energy (Commonwealth Fusion: $863M).
$20B non-exclusive technology licence plus asset purchase and acqui-hire (Dec 2025); NVIDIA states it did not acquire Groq. NVIDIA hired founder Jonathan Ross, president Sunny Madra and engineers, and CNBC reported ~$20B in cash for assets; Groq continues independently under CEO Simon Edwards and GroqCloud is uninterrupted. The message is still consolidation — and Senators Warren and Blumenthal are probing whether the structure evaded merger review.
AWS: $128.7B revenue in 2025 (grew 24% in Q4). Advertising: $68.6B in 2025 (grew 22%). Combined: $197B of AI-exposed revenue. Amazon's ~$200B in 2026 CapEx is described as covering AI infrastructure and robotics as "seminal opportunities."
If AI agents shop for users, they optimize for price/fit/speed — not for Amazon's sponsored placements. Amazon's $17.3B/quarter ad business is directly threatened. Amazon sent legal threats to Perplexity over agentic shopping, and is updating site code to deter outside AI agents.
Rufus AI shopping assistant: 250M+ users, 60%+ higher conversion. Buy for Me: agentic purchasing across 500K products on other brands' sites. 1M+ robots across 300+ facilities. Strategy: internalize agentic commerce inside Amazon-controlled rails.
The market is not forming neat vertical empires. It is forming unstable coalitions. OpenAI is loosening Microsoft dependency and moving multi-cloud, but that also means less protection from a single dominant patron. Anthropic looks more diversified: Amazon for primary training and Bedrock distribution, Google/Broadcom for TPU capacity, Microsoft/NVIDIA for Azure capacity, and now SpaceX for a near-term NVIDIA GPU burst (more than 300MW and over 220,000 NVIDIA GPUs, per Anthropic's own framing). But SpaceX/xAI is also circling Cursor, one of the coding distribution points that helped make Claude economically powerful — Cursor and GitHub Copilot together represented roughly 30% of an earlier Anthropic revenue milestone, per reporting.
The strategic question is no longer just who has the best model. It is who depends on whom for compute, distribution, revenue, and default workflow.
OpenAI / Microsoft / Oracle / AWS — exclusivity reset, multi-cloud assembly, no single protector.
Anthropic / Amazon / Google / SpaceX / Cursor — many patrons, many distribution points, and a competitor already inside the tent.
Meta — own data, own distribution, own CapEx, own silicon roadmap. Separate game from the coalition stacks.
Anthropic gets SpaceX compute. SpaceX monetizes Colossus. Cursor gets compute leverage. xAI gets a path into developer distribution. Everyone gets stronger, and everyone gets more exposed.
The stakes rise fast when intelligence gains a body. At that point, this is no longer just a product cycle; it is national strategy.
Two things matter for a portfolio rather than for a policy audience: Article 50 is live now and binds deployers, not just labs, and the US pre-release review channel has already been measured in shipped-product downtime.
AI is no longer just a market contest. It is becoming a contest between national systems, supply chains, and spheres of influence. Two rival stacks are taking shape as export controls tighten, sovereign capital accelerates, and strategic autonomy becomes a requirement.
Released Jan 20, 2025 under MIT license. Matched OpenAI o1 on AIME 2024 (79.8% vs 79.2%), MATH-500 (97.3% vs 96.4%). 20–50× cheaper to use.
Broadcom -17.4%, Oracle -13.8%, Marvell -19.1%. Philadelphia semiconductor index -9.2%.
The lesson: DeepSeek did not prove compute no longer matters. It proved that efficiency gains, better architectures, and RL-heavy post-training can narrow the frontier with much less capital than the market had assumed.
R1 scored 79.8% on AIME 2024 (OpenAI o1: 79.2%), 97.3% on MATH-500 (o1: 96.4%), 71.5% on GPQA Diamond (o1: 75.7%), and 2029 Codeforces rating (o1: 2061). Genuinely frontier-adjacent on all key reasoning benchmarks.
Alibaba released Qwen 2.5-Max on Jan 29, 2025, claiming it beat GPT-4o, DeepSeek-V3, and Llama 3.1-405B across the board. Baidu announced Ernie 4.5 would go open-source from Jun 30, 2025 — a direct strategic reversal linked to DeepSeek pressure.
By Sept 2025, the Qwen family comprised 300+ generative AI models, 600M+ downloads, and 170,000+ derivative models globally. The "China open model" wave became real at ecosystem scale.
On January 13, 2025, the US unveiled an AI diffusion rule that split the world into three tiers based on access to advanced AI chips.
Fixed allocation of 49,901 H100-equivalent GPUs through 2027, with a smaller no-license window of 1,699 H100-equivalents.
Initial chip export controls to China
Controls tightened — NVIDIA H100 restricted
DeepSeek proves constraints accelerate innovation
Three-tier diffusion framework — world split into allies, capped, blocked
Bifurcated AI ecosystem solidifying — two dominant stacks plus contested middle
The Pentagon has cleared a small set of commercial vendors to deploy frontier AI on classified networks via its GenAI.mil platform. We no longer print the vendor list, the impact levels or the personnel count: none could be verified against a DoD or CDAO release. Vendors agreed to the Pentagon’s “all lawful use” standard; the list is explicitly framed as a diversity-of-supply move that includes open-source weights alongside proprietary models. Brookings puts DoD’s potential AI contract value at $90.7B in 2026, about 98.9% of all federal AI spending. Anthropic was excluded amid an ongoing supply-chain dispute. Net effect: the geopolitical map of the AI stack now sits inside the procurement diagrams of the world’s largest customer for compute. Safety posture and supply-chain stance are go-to-market constraints, not just brand choices.
Source: Department of War · DefenseScoop · BrookingsThe federal government wants early visibility into covered frontier models, but not a licensing regime. The June 2 order creates a voluntary framework for pre-release access — up to 30 days of federal access to covered models for trusted-partner collaboration — while explicitly rejecting mandatory model licensing, preclearance, or permitting. The distinction that matters: this is voluntary review, not mandatory approval, so it lowers regulatory uncertainty for US labs without creating a gate they must pass through to ship.
Source: White HouseRisk tiers established: unacceptable (banned), high risk (strict obligations), transparency/limited risk (disclosure), minimal (no rules).
Banned AI practices, AI system definition, and AI literacy obligations start applying.
General-purpose AI obligations become applicable. Member States designate national authorities.
High-risk AI (Annex III) and transparency rules (Article 50) start applying. Formal enforcement begins.
High-risk AI embedded in regulated products gets extended transition. All categories fully enforceable.
Penalty scale: Prohibited-practice violations can reach €35M or 7% of global annual turnover — large enough to matter for any company deploying AI in Europe.
ByteDance's Doubao exceeded 100M DAU by Feb 2026, processing 1.9B AI queries during CCTV's Spring Festival Gala. Doubao-1.5-pro (Jan 2025) was priced at only 3–4% of GPT-4.
Alibaba's Qwen 3.5 (Feb 2026): claimed 60% cheaper to run, 8× larger workloads. Baidu's MuseSteamer (Jul 2025) for enterprise video generation. Tencent's Hunyuan + Yuanbao for text, image, and video.
China's Interim Measures for Generative AI Services took effect Aug 2023. By Mar 2025: 346 services filed. By end-2025: 748 generative AI services completed filing and 435 applications registered. Providers must display model name and filing number.
By 2024: 2,027,000 industrial robots in operation, 295K annual installations, 54% of global demand. Chinese manufacturers captured 57% of the home market for the first time. This manufacturing base gives China dense industrial data and a path to embodied AI scale.
The world looks less like a clean Cold War binary and more like two dominant stacks plus a contested middle:
Privileged access to frontier compute. NVIDIA + hyperscalers. Closed-weight leaders at the top, open-weight ecosystem underneath.
Lower-cost/open models, industrial AI, mass consumer distribution. H800-constrained but architecturally innovative.
Tier 2 states bargain for room to maneuver. Export controls force countries to choose alignment.
The UAE example: G42 divested China investments and removed Chinese hardware to stay inside US rules after Microsoft's $1.5B investment (Apr 2024).
Source: CSIS, Reuters, New Lines Institute>90% of advanced AI chips manufactured on a single island
Source: SIA, 2025The race for AI supremacy may not be won by who builds the best model — but by who solves their bottleneck first. The US has the chips but is hitting energy bottlenecks. China has the electrons but is constrained on silicon. Everyone else is choosing sides.
Geopolitics is accelerating capital deployment, not cooling it. That creates real opportunity, and very real mispricing.
Our conclusion has not changed: this is not a bubble. What changed is that the repricing now has a mechanism instead of a mood — the private-to-public transition has begun, and it is marking the private book in public. We keep this section to four things: public marks, multiples, funding mix and concentration.
Deliberately excluded from this section: nuclear announced-vs-delivered gaps, humanoid funding-versus-units, and Tesla robotaxi flatness. All three are real and sourced — they live in the energy, physical-AI and vertical sections respectively. Pulling them in here is what turns a calibrated section into a bear slide, and we have not added one.
Sources: CNBC (Cerebras) · Data Center Dynamics · Cerebras Q1 2026 · Adaptation.ai (Sierra ARR) · Decagon (Series D) · Reuters (Cursor) · FactSet · Ionic · Crunchbase via AI Weekly · PitchBook · Fierce Healthcare · NVIDIA Q1 FY2027 · CNBC (revenue signal)These aren’t Silicon Valley startups. This is a roofer in Texas.
Click “The Data” to see what actually happened →
“If AI were killing tech jobs, recruiters wouldn’t be the hottest hire in the market.”
Source: Lenny Rachitsky / TrueUp, 9,000+ companies, March 2026| Company | Valuation | Revenue | Multiple |
|---|---|---|---|
| OpenAI | $852B | $25B ARR | ~34x |
| Anthropic | $965B | $47B run-rate | ~21x |
| Perplexity DVC | ~$22.6B | ~$450–500M | ~45–50x |
| Cursor | $60B | $4B | ~15x |
| xAI pre-merger | $250B | ~$500M ARR | ~500x |
Pricing power depends on model scarcity — which is eroding fast
Distribution advantages can evaporate when models improve
Robotics, bio, energy — capital-intensive, long payoff horizons
| Dot-Com (2000) | AI (2026) | |
|---|---|---|
| Revenue | Minimal / speculative | $25B+ ARR at frontier |
| Enterprise Adoption | Early experiments | 53% of enterprises deploying |
| Infra Spend | Telco capex bubble | $700B+ hyperscaler capex |
| Moats | Weak — eyeballs only | Data, compute, distribution |
The answer isn't irrational exuberance — it's structural. Global AUM is heavily concentrated in a small number of very large managers who cannot write $5M seed checks, so capital funnels into the few AI companies big enough to absorb it. We no longer print the specific concentration statistic we previously carried: it had no source.
Top 10 manage $62T — 42% of global AUM. They need to deploy billions per quarter. Early-stage VC is invisible to them.
Source: BlackRock Q4 2025 earnings, Vanguard, UBS, Fidelity filingsAI captured roughly 61% of all global VC in 2025. The denominator is reported at both $427B and $440B depending on the tracker cut, so we give the share rather than assert one total. But within AI, concentration is extreme:
Tap any bar for details. These 8 rounds alone total ~$250B+ — more than total global VC outside AI.
Source: CNBC, Crunchbase, OECD, company announcements 2025–2026The vast majority of investors can't write small, high-risk VC checks. They need to deploy at scale. So they pile into the ~20 private AI companies large enough to absorb $100M+ checks. Insane competition for a tiny number of deployable deals drives late-stage multiples up — which are then inherited by early-stage VCs.
Sovereign Wealth Funds deployed $46B into AI ventures in 2025. Saudi PIF, Abu Dhabi's Mubadala, Singapore's GIC and Temasek are now among the largest AI investors.
Source: EY Global GenAI VC Report, Dec 202597% of AI deal value went to North America. San Francisco Bay Area alone captured $122B — 75%+ of all US AI funding.
Source: OECD, Crunchbase 2025 dataBe skeptical on valuation. Not on AI demand. The top quartile is pulling away. The bottom quartile is standing still. The gap widens every quarter.
Noise in pricing does not change the direction of the market. It makes disciplined positioning across the stack even more valuable.
The entire world's business is being rebuilt. Every industry — search, commerce, legal, healthcare, finance, property, video, code, robotics — is getting a new AI-native entrant. The incumbents have resources, but they can't do everything at once. They have organizational drag, legacy architectures, and cannibalization risk that slows them down.
The opportunity for startups isn't to avoid competing with Google and OpenAI. It's to move faster in the seams they can't fill. Perplexity competes with Google Search. Cursor competes with GitHub Copilot. Higgsfield competes with Sora. In every case, the startup has the advantage of focus, speed, and a willingness to bet the company on a single wedge.
DVC's portfolio is positioned across the stack because the winners won't cluster in one layer — they'll emerge wherever a startup can claim turf faster than an incumbent can defend it.
A portfolio is a point of view made concrete. From here, the only question that matters is where the forces already in motion take us.
These are not predictions pulled from thin air. They are the logical consequences of every trend in this presentation — extrapolated five years forward. The pattern is clear: AI software moves fast, AI hardware moves slower, and AI adoption in large legacy industries moves slowest of all.
"The 2026–2030 AI story is less about whether AI works, and more about whether institutions — constrained by power, regulation, and labor — can absorb ultra-cheap intelligence quickly enough. The winners will have culture that retains talent, distribution that reaches users, and the discipline to invest across the full stack rather than betting on a single layer."
The trends are clear. The question is who acts on them first.
You are not late to a trend.
You are early to a restructuring of the global economy.
The reset is not to learn more AI jargon. It is to see the stack clearly, understand where value is shifting, and act before today's temporary leaders harden into tomorrow's incumbents.
That is where DVC operates: across the system, across the cycle, and with a bias toward the layers that compound as AI moves from breakthrough to infrastructure.
August 2026