← Full Report
1 / 18

DVC presents

STATE OF AI

The Operating Manual for the AI Revolution

A 20-minute compressed walk through the full stack — where money flows, who wins, what changes.

Read the full report state-of-ai.dvc.ai Scan or visit · August 2026
Press → or click to begin
→ Full report

This runs today.

A day in the life — already automated

→ Dive deeper

Tech cycles

The price of intelligence is collapsing

Every platform made one scarce input cheap. AI makes custom software cheap enough to give every workflow its own brain.

2023 was ignition. · 2026 is still the beginning.

Same software arc. · New cost curve. · Very beginning.

→ Dive deeper

The Stack · 1 of 3

Five layers. One economy.

$60B at the top generates ~$725B of buildout below (Big 4 2026 midpoint; up to ~$745B at top end).

Revenue vs Margin · 2 of 3

Same stack. Two lenses.

Revenue grew everywhere. Margins didn't follow. The top earns software economics. The bottom earns utility.

Forces on the Stack · 3 of 3

Two forces. One stack.

Down: commoditization. Up: personalization. The moat moved up.

$60Bapps revenue · ~$725B2026 capex (mid · up to ~$745B) · 12 : 1infra : app revenue
+400%app revenue growth · 50%app gross margin · 12%energy margin (utility)
10×spread between commodity and frontier list prices, Aug 2026 · moat ↑value migrates up · $60B → ~$725B12 : 1 stack ratio

Tap a state to switch lenses

→ Dive deeper

Application layer

The user never sees the model.

Whoever owns the application picks the model. Distribution > intelligence.

LAB LEADERSHIP · AUG 3, 2026 Private valuations are primary-round marks; secondary prints differ. Every private ARR figure is company-disclosed and unaudited.
$965B ANTHROPIC — primary; $65B Series H closed May 28 2026. $47B run-rate reported June 2026, up from the official $30B in April. 42.4% of US paid business use vs OpenAI 39.5%. 300K+ business customers, ~70% of the F100 (company-disclosed).
$852B OPENAI — primary; ~$880–895B implied by secondary prints. ChatGPT mobile apps crossed 1B MAU (May 2026, Sensor Tower estimate); 900M WAU is OpenAI’s own last disclosure, Feb 2026. Codex + ChatGPT Work at ~10M WAU (OpenAI-disclosed; active undefined). No audited 2026 revenue filed.
950M GOOGLE GEMINI MAU; AI Mode past 1B MAU; ~90% of the F100 on Gemini Enterprise; 22B tokens/min through the APIs.
1.2B META AI MAU / 800M WAU — nominally the largest assistant by reach, but embedded in existing search bars rather than chosen.

Sources: Reuters · Ramp AI Index, July 2026 · The Verge · OpenAI (usage) · CNBC (revenue signal)

→ Dive deeper

Model layer · August 3, 2026 flagship set

The model market became a barbell.

Premium multi-day agents priced up to $10/$50. Capable output priced down to $1–$6. The middle thinned — and token efficiency became a third, independent economic axis.

$10 / $50 PREMIUM MULTI-DAY AGENTIC (Claude Fable 5 / Mythos 5)
$1 – $6 CAPABLE OUTPUT (GPT-5.6 Luna, Sonnet 5, Muse Spark)
PREMIUM AGENTIC Priced for multi-day autonomous work, not chat
Claude Fable 5 / Mythos 5$10/$50
Jun 9 · multi-day agentic tier. SWE-Bench Pro leader at 80.3%. Mythos 5 is restricted to Project Glasswing partners, not generally purchasable.
GPT-5.6 Sol$5/$30
Jul 9 GA · Ultra mode, Max reasoning. 54% more token-efficient on coding than its predecessor.
The top of the market got more expensive, not cheaper.
EFFICIENT FRONTIER Frontier-adjacent capability at commodity prices
Claude Sonnet 5$2/$10
Promo through Aug 31, then $3/$15. 1M context. Default for Free/Pro and Claude Code.
Grok 4.5efficient frontier
Strategically important; exact price, context and token comparison removed after primary-source audit.
GPT-5.6 Terra / Luna$2.50/$15 · $1/$6
Luna at $1/$6 is the floor of the credible frontier tier.
Gemini 3.6 Flashefficient frontier
Lower-output-token positioning; exact pricing, percentage and cutoff removed after source audit.
Meta Muse Spark 1.1$1.25/$4.25
Meta's first paid Model API. Leads on tool use; loses on raw coding and long context.
OPEN-WEIGHT CHINA The parameter-count crown left the West
Kimi K3 (Moonshot)2.8T
896 experts / 16 active · 1M context · weights released Jul 27. Largest open-source model to date.
DeepSeek V4 Previewopen weights
Parameter and active-expert splits removed because they could not be confirmed from a primary release.
Open weights now win on scale and ownability — not on price.
RESTRICTED / CYBER Cyber capability became a release gate, not a footnote
Fable 5 / Mythos 5 controlsrestricted access
The exact suspension duration and regulatory causality were removed because they could not be independently verified.
GPT-5.6 pre-release reviewEO 14409
A US pre-release review process preceded GA. The White House denied granting a “green light”; OpenAI objected to it becoming “the long-term default.”
Gemini 3.5 Flash CyberDedicated SKU
Google shipped a cyber-specific model; GPT-5.6 was billed as OpenAI's “strongest cybersecurity model yet.”
Benchmark integrity: the Grok 4.5 CursorBench result was withdrawn after Cursor disclosed that a snapshot of its codebase had accidentally entered the model's training data. Treat single-benchmark leadership claims accordingly.

Sources: Anthropic · OpenAI GPT-5.6 · The Decoder · Meta AI · How AI Works · Elser.ai (China) · Roo (CursorBench withdrawal)

→ Dive deeper

Capability vs. cost

Smarter AND cheaper.

Every prior cycle: more power = more cost. AI inverted that. Embed intelligence everywhere.

THREE SEPARATE VENDOR CLAIMS, NOT A TIME SERIES Each figure is a single vendor's own comparison against its own predecessor. They are not measured on a common benchmark and must not be read as a curve.
VENDOR CLAIM Grok 4.5 was marketed on lower tokens per task; the exact multiplier was removed after source audit.
54% GPT-5.6 Sol is 54% more token-efficient on coding than its predecessor.
VENDOR CLAIM Gemini 3.6 Flash was marketed on lower output-token use; exact percentage and price delta were removed.
NEW ECONOMIC AXIS Price-per-token understates the decline in cost per completed task. Application-layer gross margin improves faster than any price curve implies.

Source for quantified efficiency claim: OpenAI. Grok and Gemini cards are retained only as qualitative vendor-positioning signals.

→ Dive deeper

Infrastructure & Energy

Software demand. Industrial buildout.

Top-four 2026 capex guidance now sits at ~$725B at the midpoint, up to ~$745B — up to ~$800B for calendar 2026 including leases, with Google raised to $195–205B. The number held. The way it is paid for did not.

$224B2024 CapEx (BASELINE)
→
~$725B2026 GUIDANCE, TOP FOUR (UP TO ~$745B)
3.1×in 2 years
HOW IT’S FUNDED The scale survived. The funding quality changed — this is now a balance-sheet story, not an operating-leverage story.
9% → 32% Incremental annual debt as a share of hyperscaler capex, FY24 → LTM mid-2026.
$0 Buybacks at Meta and Alphabet, 2026: $0 — below the $12.6B Q4 2025 trough. Per-company split not printed — unsourced.
FCF negative Alphabet free cash flow turned negative for the first time since its 2004 IPO.
$84.75Bprogram size, not cash collected Alphabet June equity raise — its first since 2005.

Hyperscaler Capital Expenditure

~$725B2026 GUIDANCE (mid · up to ~$745B)
Show AI Revenue vs CapEx gap
Source: SEC filings, company earnings, Goldman Sachs, Epoch AI — 2026 figures are guidance midpoints
→ Dive deeper

Silicon & Power · the two physical gates

The monopoly cracked. The power didn't arrive.

Three independent breaks in NVIDIA's accelerator lock happened in one quarter. Meanwhile announced nuclear runs 5× ahead of operational nuclear — and gas is filling the gap.

SILICON The accelerator monopoly cracked in one quarter
NVIDIA Q1 FY27
$81.6B quarterly revenue · data center $75.2B

VeraRubin production shipments are planned for Q3. The physical-AI TTM and China-revenue figures previously shown here were removed because they could not be reproduced from the first-party release.

AMD hyperscaler commitments
INSTINCT major hyperscaler commitments

AMD’s roadmap won material hyperscaler adoption. Exact gigawatt totals were removed because the cited report could not be independently reproduced during the audit.

BROADCOM · CUSTOM +143%
$10.8B AI semis FQ2 · Q3 guided $16B

OpenAI partnered with Broadcom on its first in-house processors. Custom silicon is no longer a side bet.

ETCHED DVC Jul 23 · Series C
$10.3B $300M led by Sequoia · >$1B raised

a16z, Jane Street and SK Hynix joined — a transformer-only ASIC bet. DVC portfolio company. Pre-order figure not printed: unsourced.

POWER Announced nuclear vs. electrons on the wire
Committed across 13 hyperscaler nuclear deals
9.8 GW
Actually operational today
1.92 GW
5:1 announced-to-delivered gap
THE BRIDGE IS GAS, NOT FISSION
2.67 GW Project Kilby — off-grid West Texas gas plant, 20-yr PPA with Microsoft. First delivery targeted 2028; final investment decision not yet taken, so this is not committed capital.
H2 2027–28 Three Mile Island / Crane — 835 MW on a 20-year PPA, ~$16B restart. Produces no power until H2 2027 at the earliest.

The binding constraints are now HBM supply, land, electricity and construction labour — not GPU fabrication.

— Jensen Huang, NVIDIA Q1 FY27

Sources: NVIDIA Q1 FY2027 · Reuters (AMD/OpenAI) · Reuters (Etched) · SMR Intel tracker, May 2026 cut · TechCrunch (Kilby)

→ Dive deeper

Agent anatomy

It's not one product. It's a new stack.

47 mapped companies across 7 layers, plus an emerging control plane above them (metrics as of Aug 4, 2026): governance, agent registries, permissions, observability, and evals. Most didn't exist 18 months ago.

1 UI / Frontend Layer How agents meet users
CopilotKit 29.3K ★ DVC Vercel AI SDK 22.6K ★ AG-UI 12.4K ★ DVC Streamlit Gradio
12.4K GitHub ★ in <1 year — AG-UI is becoming the standard event protocol for agent-to-user interaction.
2 Orchestration Planning, routing, multi-agent
LangGraph 129K ★ CrewAI AutoGen Semantic Kernel Pydantic AI Sixtyfour Kapso
ReAct loops → multi-agent systems with planners, workers, verifiers. Most serious startups eventually build proprietary orchestration.
3 Memory State beyond the prompt window
mem0 49.6K ★ DVC Letta / MemGPT Zep Kite
+26% accuracy over OpenAI Memory, 90% less tokens. mem0 externalizes memory — works with any stack.
4 Tool / Action How agents act on the world
MCP A2A 22.5K ★ Function Calling Browserbase Composio Firecrawl Browser Use Exa Orthogonal
"MCP is becoming the REST of the AI era" — MCP for tools, A2A for agent-to-agent, AG-UI for agent-to-user.
5 Foundation Models The reasoning engine(s)
OpenAI Anthropic Google xAI Meta / Llama DeepSeek Mistral
37% of enterprises use 5+ models in production. Multi-model routing is standard — Harvey uses 6+ providers.
6 Execution Where code actually runs
E2B Daytona WebContainers Modal Dynamo NVIDIA OSS
3 patterns: local, cloud sandbox, browser-native. Cursor = local, Devin = cloud, Bolt = browser. Dynamo: 30× inference throughput.
7 Eval, Voice & Comms Observability, quality, output
ElevenLabs Vapi LangSmith Phoenix / Arize Braintrust AgentMail
Best startups treat eval as product, not afterthought. Harvey: BigLaw Bench. Perplexity: search_evals.
Sources: Crunchbase, TechCrunch, GitHub, npm, PyPI · 2025–2026
→ Dive deeper

Vibe coding

Code is becoming a throwaway artifact.

42% of committed code is AI-generated or significantly AI-assisted. Cursor reached $4B ARR and was acquired for $60B all-stock. Lovable hit $500M ARR on 146 employees — ~$3.4M ARR per head.

CATEGORY CONSOLIDATION Independent coding tools now face acquisition-consolidation and platform-bundling at the same time. This is model + application vertical integration, not a valuation event.
Independent tool ↓ Cursor → SpaceXAI $4B ARR; $60B all-stock agreement signed Jun 16, subject to closing. The IDE and frontier-model ambitions now sit inside one public-company structure.
Model-layer surface ↓ Claude Code → Sonnet 5 bundle Claude Code became the default surface for Sonnet 5 from Jul 1, at a promo $2/$10. The coding tool ships with the model.
Assistant surface ↓ Work + Codex → OpenAI platform ChatGPT Work shipped Jul 9; Codex + Work reached ~10M WAU by Jul 21, doubling in nine days (OpenAI-disclosed; “active” undefined).
New entrant ↓ Muse Spark → Meta enters Meta entered coding directly on Jul 9 with Muse Spark 1.1 at $1.25/$4.25 — its first paid Model API.

Margin note: Cursor’s reported gross margin is close to negative on token pass-through, which makes a ~15× LTM ARR acquisition multiple the expected outcome for this cost structure, not a discount.

Source: SonarSource 2026 State of Code Developer Survey (42% AI-authored code, 72% daily use among triers). Stack Overflow 2025 — 51% daily — as secondary signal.

→ Dive deeper

Voice AI · the category that cleared the bar

Voice stopped being a feature. It became a layer.

A year ago voice was a demo. In July 2026 it has a $600M ARR independent, a full-duplex frontier model, and 41% of the Fortune 500 paying for it.

ELEVENLABS Independent · July 2026
$600M ARR — from $330M at end-2025 (~175% YoY)
  • 41% of the Fortune 500 as customers
  • 1B+ end users reached via API
  • $22M paid out to 10.4K+ voice creators
Company-disclosed, unaudited (CEO, All-In conference)
OPENAI GPT-LIVE 2026 product release
Full duplex OpenAI introduced a full-duplex voice interface
  • Simultaneous listen-and-speak — no push-to-talk turn model
  • Platform-level: voice ships with the frontier model, not beside it
  • The independent layer and the platform layer now overlap
Primary source: OpenAI product announcement
$600M ELEVENLABS ARR
41% OF THE FORTUNE 500
1B+ API END USERS
FleetWorks · Avoca DVC VOICE-AGENT EXPOSURE

Sources: Postbeam and Bleap (ElevenLabs, company-disclosed) · OpenAI GPT-Live

→ Dive deeper

Business models · usage is the engine

Outcome pricing stopped being a thesis. It became a price list.

In one quarter the category acquired published prices, an incumbent convert, a quantified TAM at risk and a services proof point. What is unresolved is margin, not mechanism.

👤
Per-seat
≡
Per-token (API)
∞
Per-subscription
◉
Per-ad
✓
Per-outcome / per-task
A PRICED MARKET Published per-outcome prices now exist across the category — not projections, list prices.
$2Salesforce Help Agent — per resolved issue; no charge on escalation
$0.10Salesforce Flex Credits — per action
OUTCOMESCognizant reports new BPO contracts moving to outcome-based models
GAAgentforce Help Agent shipped pay-per-resolution GA in July — and Salesforce closed the m3ter acquisition Jul 1 to meter and bill it.
$234BGartner: ~20% of enterprise application software spend is exposed by 2030 to “agentic arbitrage” — outcomes delivered without the seat login.
45%of Cognizant’s new BPO contracts are now signed on outcome-based commercial models. Services pricing is converting too.

Charge for the work. Anthropic: $47B reported run-rate (June 2026), up from the official $30B in April. Perplexity: >$450M ARR after Computer (Mar 2026, company-disclosed). OpenAI: 900M WAU was the last company-disclosed February figure. ARPU and paid-conversion comparisons use modelled, non-comparable user bases and are shown only as directional context.

ENTERPRISEEnterprise-heavy model usage monetizes differently from consumer reach. Existing ARPU estimates mix incompatible MAU and WAU denominators, so exact comparisons were removed.
HYBRIDSeats, usage and outcomes coexist. A subscription floor still absorbs volatility while the meter moves toward completed work.
MARGINThe unresolved question is not whether outcome pricing exists, but whether the price covers inference, escalation and implementation.
Sources: CIO / Gartner · Teqfocus (Salesforce) · Moneycontrol (Cognizant)
→ Dive deeper

Services disruption

Where AI autopilots are attacking services.

Vertical market sizes for ten categories after Sequoia Capital (Bek, March 2026). Remaining categories, quadrant grouping and company selections are DVC’s. Funding figures checked Aug 4, 2026.

→ Dive deeper

How founders attack the services map

Three angles. Pick one.

SaaS to incumbents. Vertical agentic. Agentic rollup. The market structure picks the strategy — not the founder.

→ Dive deeper

Healthcare AI · incentive mechanics

AI makes healthcare more efficient — and more expensive.

AI follows incentives. In fee-for-service, the same visit documented better is a higher-severity claim. Efficiency and total spend both go up.

THE MECHANISM
WORKFLOW TIME Independent studies show ambient documentation can reduce documentation burden; effect sizes vary by workflow and adoption.
CLAIM SEVERITY & COST PwC's mechanism is not more services — it is “changes in severity, mix and amount per claim.”
WHAT THE PAYERS NOW SAY
9% 2027 medical cost trend, group plans (8.5% individual) — the highest in nearly two decades
~70% of health plans rank provider AI documentation and coding in their top three cost inflators. ~20% call it the single biggest.
SEVERITY Payers identify documentation and coding intensity as a mechanism that can raise the amount paid per claim.
TWO GATES
01 REIMBURSEMENT decides what gets bought
  • WISeR brought AI-assisted prior authorization into Original Medicare and triggered congressional resistance.
  • ePA becomes mandatory Jan 1, 2027 for MA, Medicaid, CHIP and FFE plans.
02 LIABILITY decides what gets deployed
  • 76.6% of harmful errors across every tested clinical AI were omissions, with potential for severe harm in up to 22.2% of cases — NOHARM, 100 real cases across 10 specialties and 31 models; not peer-reviewed, disputed.
  • OpenAI sued for the unauthorized practice of medicine one day before Health in ChatGPT launched.
TWO CLOCKS, NOT ONE
2–4 yrs ADMIN — moves on regulatory deadlines (ePA 2027, NHS App triage by April 2028)
10–15 yrs CLINICAL AUTONOMY — moves on evidence, liability and regulated deployment rather than vendor pilots.
DVC PORTFOLIO
DOCTRONIC Acquired Summer Health in July — adding pediatric domain data and expertise.
COLLECTLY Live in the Epic Showroom with bidirectional Epic integration.company announcement

Sources: JAMA Network (ambient-documentation direction) · Axios and Fortune (PwC 2027 trend) · STAT (WISeR) · NOHARM abstract · Reuters (lawsuit)

→ Dive deeper

Physical AI

It already left the screen.

$38 trillion labor market in play. The Waymo–Tesla gap widened rather than closed, and humanoids reached the public markets before the units did.

FUNDING RAN AHEAD OF DELIVERY Humanoid funding hit $8.6B in 2026 by late July — 1.8× all of 2025 — with two public-market events. The delivery ledger has not moved with the cap table.
$39B FIGURE AI reported valuation on $2.34B raised — the sector’s high-water mark.
~$2.5B AGILITY ROBOTICS going public via SPAC, raising $620M+ gross expected — the SPAC is announced, not closed — and its CEO declined to promise a home robot.
PUBLIC-MARKET PLANS Several humanoid companies announced large rounds or public-market plans, but company-specific valuation and IPO-target figures were removed when they could not be independently confirmed.
THE CAVEAT Large humanoid rounds have not yet produced matching unit deliveries. NVIDIA now reports physical AI as a meaningful infrastructure line, but the trailing-twelve-month revenue figure previously printed here could not be independently confirmed.

Sources: RoboFutur and CNBC (Waymo metros) (Waymo) · Reuters and Electrek (Tesla) · Venture Post and GrabaRobot (humanoids). Rhoda figures per DVC.

→ Dive deeper

Key learnings

Six things we now believe.

Revised August 4, 2026. Where value is shifting, what's structural, and which of our own beliefs the last quarter forced us to retire.

  1. 01

    Models commoditize. Distribution wins — but distribution changed meaning.

    The contested surfaces stopped being app stores and became records, regulators and institutional accounts. OpenAI now connects hospital records and Apple Health inside ChatGPT; Mayo Clinic owns a frontier clinical model that Microsoft distributes; OpenAI and Anthropic donated public-health seats rather than sell them. Leadership also inverted: Anthropic at 42.4% of US paid business use vs OpenAI’s 39.5% (Ramp AI Index, July 2026).

  2. 02

    Energy is still binding — and it is being met with gas, not fission.

    9.8 GW of hyperscaler nuclear committed against 1.92 GW operational in the May 2026 tracker snapshot. Project Kilby is a proposed 2.67 GW off-grid gas bridge, but it had no final investment decision as of Aug 4, 2026. The real constraints are construction and financing: debt went from 9% to 32% of hyperscaler capex and Alphabet's free cash flow turned negative for the first time since 2004.

  3. 03

    Outcome pricing has been invented. Margin is the unresolved variable.

    We are retiring “the pricing model is still being invented.” Salesforce now lists per-resolution and per-action pricing, and Cognizant reports new BPO contracts moving toward outcome-based commercial models. Gartner puts $234B (~20% of enterprise app spend) in scope by 2030. What remains unresolved is whether those unit economics survive token costs.

  4. 04

    Open weights lead on scale. Regulated inference stays proprietary.

    Chinese open weights took the parameter crown through Kimi K3 at 2.8T, according to Moonshot’s own release. They matter because they are ownable, not simply because they are cheaper. Unsupported DeepSeek parameter splits and secondary-source price comparisons were removed. Regulated enterprise inference remains proprietary or hybrid.

  5. 05

    Coalitions, not empires.

    Three stacks now define the frontier. OpenAI / Microsoft / Oracle / AWS: exclusivity reset, multi-cloud assembly, no single protector. Anthropic / Amazon / Google / SpaceX / Cursor: diversified compute and distribution, with a competitor (xAI) already inside the tent. Meta: separate, vertically controlled. Everyone gets stronger; everyone gets more exposed.

  6. 06

    AI follows incentives. Efficiency doesn’t guarantee lower spend.

    Health plans project a 9% 2027 medical cost trend — the highest in ~two decades — and ~70% rank provider AI documentation and coding among their top three inflators. Same care, higher-severity code, no rise in denials. Read across: in any fee-for-service market, AI raises throughput and the bill. Humanoids stay “real but slower” for the same class of reason — $8.6B of 2026 funding against a delivery ledger that hasn’t moved.

Not a bubble — a repricing with a mechanism. Public markets have started marking the private book, and concentration (OpenAI + Anthropic took 43% of every H1 2026 venture dollar) is now market structure rather than a warning. You are not late to a trend; you are early to a restructuring.

→ Dive deeper
← → Space · Esc to exit