moatwatch

an mcgrathlabs observatory

Tracking the commoditization of intelligence: how fast open-weight models close on the frontier, what a token actually costs — API, hosted open, or owned hardware — and pre-registered turning points so the regime change is visible before it’s a headline.

As of 2026-09-18

Open weights are currently
behind the closed frontier (GLM 5.3 Flash vs Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback))
16 frontier releases not yet benchmarkedthe frontier reference is itself subject to policyhow current is this number? →
Open share of tokens
72%
OpenRouter daily volume · 88% of it Chinese-published
Hosted open floor
$0.14
per Mtok, near-frontier open
Frontier list price
$20.00
per Mtok, best closed (blended 3:1)
Local prosumer cost
$8.14
per Mtok @60% utilization

The thesis

Frontier intelligence is losing its scarcity value. Open-weight models now arrive months — not years — behind the frontier, already carry most measured API traffic, and every serving price is falling toward marginal cost. Meanwhile the five hyperscalers are spending ~$700B a year — nearly all their operating cash flow — on compute whose return depends on that scarcity holding. Something gives: either AI revenue materializes fast enough to justify the buildout, or capex breaks and takes the semiconductor complex with it.

moatwatch pre-registers what “giving” looks like — the day open weights reach the frontier, the day owning beats renting, the day the price war goes vertical, the day spenders and suppliers reprice together — and checks every day whether it happened. The goal is to see the turning points in the data before they become headlines. Four questions follow, in order: is the gap closing, is the price collapsing, does open stay open, and can the buildout be paid for.

The board

2 tripped · 1 warming · 12 pre-registered

Every condition below was written down before it could trip, so a trip is a signal and not a narrative. They do not all mean the same thing — read the groups.

Commoditization confirmed

1 tripped · 0 warming

Capability and price converging. A trip here means intelligence became a commodity on schedule or sooner.

Watching22 % gap

Open weights reach the frontier

Best open-weight model comes within 5% of the best closed model on the Artificial Analysis Intelligence Index. Intelligence is commoditized one generation sooner than expected.

Trip condition: gapPct ≤ 5 (warming ≤ 10)

TRIPPED72 % of tokens

Open weights carry most real traffic

Open-weight models exceed 50% of OpenRouter daily token volume. Revealed preference — the market votes with paid tokens, not benchmarks.

Trip condition: openTokenShare ≥ 50 (warming ≥ 40) · tripped 2026-03-19

Watching57 × hosted floor

Owning beats renting at realistic utilization

A sub-$10k prosumer rig at 60% utilization produces tokens cheaper than the cheapest hosted near-frontier open model. The moment self-hosting stops being a fantasy for small teams.

Trip condition: prosumer $/Mtok @60% util ≤ hosted open floor (warming ≤ 1.5×)

Watching-82 % 90d decline

Frontier price war goes vertical

Blended $ per intelligence-index point of the best closed model falls 50%+ within 90 days. The labs are torching margin to defend share — commoditization priced in by the producers themselves.

Trip condition: frontierPricePerIntel drops ≥50% over trailing 90d (warming ≥ 30%)

Watchingno data yet

Cheaper tokens stop paying for themselves

Implied daily inference spend (top-50 token volume × the hosted open price floor) falls 20%+ over 90 days. The Jevons assumption underwriting the entire buildout — cheaper intelligence pulls in more than enough extra demand — failing in the measured data. The bridge between falling prices and the capex bill.

Trip condition: 90d change in volume × price ≤ −20% (warming ≤ 0%)

Thesis fragility

1 tripped · 1 warming

The thesis breaking for reasons unrelated to capability — an open ecosystem one government can withdraw, weights that stop being downloadable, or a buildout that blinks first.

TRIPPED88 % of open tokens

The open majority is single-jurisdiction

Chinese-published models serve 80%+ of open-weight token traffic. Not a commoditization signal — a fragility one: an open ecosystem this concentrated can be withdrawn by one government's export-control decision, which is exactly what Beijing is reported to be weighing. The higher this runs, the less the open majority proves about durable commoditization.

Trip condition: CN share of open-weight tokens ≥ 80% (warming ≥ 65%) · tripped 2026-07-24

Watching0.0 % of open traffic

Open weights stop being downloadable

Models serving 10%+ of open-weight traffic have weights that are gated or have been withdrawn — verified daily against the publishers' own repositories, not inferred from the creator's name. The mechanism that would end the thesis without any change in the technology: an open majority that can no longer be downloaded is an API business with extra steps. Counts only real restrictions; a new release whose weights are not up yet is tracked separately as unverified.

Trip condition: gated or withdrawn ≥ 10% of open traffic (warming ≥ 3%)

Warming0.0 in force

A government restricts open weights

Any major jurisdiction puts a restriction on open-weight model access into force — Beijing limiting foreign weight downloads, Washington restricting enterprise use of Chinese open models, or the equivalent. Warming while such a measure is credibly reported or formally proposed. Hand-curated from primary reporting in data/policy-events.json; reported intentions never count as enacted rules.

Trip condition: enacted open-weight restrictions ≥ 1 (warming: ≥1 reported/proposed)

Watching0.0 guidance cuts

A hyperscaler blinks on capex

Any top-5 hyperscaler guides capital expenditure down year-over-year. The single loudest 'something gave' signal available — tracked manually from earnings guidance in fundamentals.json.

Trip condition: guidance cuts YoY ≥ 1 (manual quarterly update)

Priced by the market

0 tripped · 0 warming

Whether equities have started to agree. Confirmation, never evidence — prices move on sentiment first.

Watching1.0 % both down

Hyperscalers and semis fall together

Both the hyperscaler basket and semis (SMH) sit 15%+ below their trailing 90-day highs at once. Not a rotation — the whole AI capex trade repricing simultaneously, the Cisco-2000 shape.

Trip condition: min(|drawdown|) of both baskets ≥ 15% (warming ≥ 10%) · tripped 2026-07-18

Watching7.2 pp spread

Spenders and suppliers decouple

Hyperscaler basket and semis 60-day returns spread beyond 15 points. The see-saw scenario: the market rewarding capex cuts (or punishing them) instead of moving both sides together.

Trip condition: |hyperscaler 60d return − SMH 60d return| ≥ 15pp (warming ≥ 10) · tripped 2026-08-19

Watching-2.4 pp spread

Market splits frontier from commodity silicon

Nvidia's 60-day return runs 20+ points ahead of the commodity-silicon basket (AMD, AVGO, MRVL, MU, TSM). Open-weight commoditization pricing into the chip complex — inference silicon commoditizes while frontier-training silicon keeps its moat.

Trip condition: NVDA 60d return − commodity basket 60d return ≥ 20pp (warming ≥ 12) · tripped 2026-08-19

Stress vs panic

Two independent composites: what the fundamentals say (thesis stress, no market inputs) vs what the market has priced (panic, prices only). The spread between them is the edge.

Thesis stress47
Market panic37
Divergence+10
Aligned

Market pricing and thesis fundamentals roughly agree. No edge either direction — keep watching.

Thesis stressMarket panic
Methodology (fixed ramps — changing one is a logged decision)

Thesis stress uses economic signals only, no market prices: frontier gap (30%→0 … 5%→100, w25) · open token share (20→0 … 60→100, w25) · 90d frontier price collapse (0→0 … 50%→100, w15) · local-parity ratio (30×→0 … 1×→100 log, w15) · capex/OCF (33→0 … 100→100, w15) · guidance cuts (any→100, w5). Missing components drop out with weights renormalized.

Market panic uses prices only: mean 90d drawdown of the hyperscaler and SMH baskets (0→0 … 25%→100, w60) · SMH 20d realized vol annualized (15%→0 … 60%→100, w40). Divergence zones at ±20.

I

The gap

Is open capability closing on the frontier?

The distance between the best open-weight model and the best closed one, on a single benchmark index. The thesis dies here if the gap stops shrinking — and the number is only as fresh as the benchmark coverage behind it.

The frontier gap

Best open-weight model vs best closed model, Artificial Analysis Intelligence Index (% behind). Shaded region is curated backfill; numbered lines are key releases; the dashed marker is the provisional nowcast, an estimate.

  1. 1. GPT-4 Frontier gap at its widest — no credible open competition.
  2. 2. Llama 2 First serious open-weight release with a commercial license.
  3. 3. Llama 3.1 405B First open model marketed as frontier-class.
  4. 4. DeepSeek R1 The 'DeepSeek moment' — $600B single-day Nvidia drawdown.
  5. 5. Kimi K2 1T-param open MoE; agentic benchmarks near closed models.
  6. 6. Price-war reports WSJ: OpenAI weighs deep price cuts to defend enterprise share vs Anthropic.
  7. 7. Kimi K3 2.8T open MoE within ~7% of frontier; SOX enters bear market territory.

How current is that number?

The gap compares the best models Artificial Analysis has already scored. Releases outrun benchmarks, and the frontier reference can be withdrawn by policy — both are qualifiers on the headline, tracked here rather than assumed away.

Provisional gap (nowcast)0.0%±1σ range -3.7% to 3.7%

If OpenRouter-relayed AA scores hold: Qwen3.8 Max (53.4 est) vs Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) (53.4, measured). Provisionally inside the 5% frontier-gap trip threshold — turning points evaluate measured data only, so this estimate cannot trip them.

Calibration: the relay matches our AA feed within ±0.5 pts on 25 of 39 overlapping models (median offset 0.0); band is ±2.0 pts (prior — tightens as nowcasts get reconciled against real scores). Track record: 2 reconciled, median abs error 6.3 pts. An estimate, not a measurement.

The frontier reference is itself subject to policy

  • 2026-06-09 Anthropic Fable and Mythos withdrawn on government security concerns (Reuters)

The gap is measured against the best available closed model. Withdrawing a frontier model narrows it for a reason that has nothing to do with open weights catching up — read the number with that in mind, and see Policy watch.

16 frontier-candidate releases await a benchmark

ReleaseWeightsReleasedWaitingRelayed index
GPT-6 Astra Pro OpenAIclosed2026-09-0414d
Gemini 3.8 Flash Googleclosed2026-09-0216d41.2
Muse Spark 1.3 Meta · 2 variantsopen2026-09-0216d
Muse Spark 1.2 Contributor Metaopen2026-08-2128d
Gemini 3.7 Flash Googleclosed2026-08-1336d39.4
Muse Glimmer 30B Metaopen2026-08-0940d
Qwen3.8 Max Qwenopen2026-08-0346d53.4
Claude Opus 5 (Fast) anthropicclosed2026-07-2456d
Gemini 3.6 Flash Googleclosed2026-07-2159d34.3
Muse Spark 1.1 Metaopen2026-07-1664d34.3
GPT-5.6 OpenAI · 3 variantsclosed2026-07-0971d
Grok 4.5 xAIclosed2026-07-0872d39.1
GLM 5.2 Z.aiopen2026-06-1694d34.0
Kimi K2.7 Code MoonshotAIopen2026-06-1298d26.3
Claude Fable 5 Anthropicclosed2026-06-09101d49.7
Qwen3.7 Qwen · 2 variantsopen2026-06-03107d25.8

Relayed indices are OpenRouter’s echo of an AA score our feed has not published yet — provisional, calibrated in the nowcast above, and never entering measured metrics or turning points. The median release→benchmark lag observed so far is 12d.

II

The price

Is the cost of intelligence collapsing, and does anyone still capture value from it?

Three prices for the same token — the frontier list price, the cheapest hosted open model, and hardware you own. Then the question that decides whether falling prices are bullish or fatal for the buildout: does volume grow faster than price falls?

The price of a token

Blended $/Mtok (3:1 input:output), log scale. Frontier list price vs the cheapest hosted near-frontier open model.

Frontier (best closed)Hosted open floor

Own vs rent

Effective $/Mtok by sourcing strategy at your volume. The turning point: when a sub-$10k box beats the hosted floor at realistic utilization.

Frontier API
best closed model, list price
$20.0/Mtok
Hosted open floor
cheapest near-frontier open API
$0.14/Mtokcheapest
Mac Studio M3 Ultra 512GB
large MoE, 4-bit (DeepSeek/K2-class)
— over rig capacity at this volume
2× RTX 5090 workstation
70B-class dense, 4-bit
$2.86/Mtok
8× H100 node
frontier-scale open MoE (K3-class), batched
$69.7/Mtok

Local cost = hardware amortized over 36 months ÷ monthly volume + electricity. Excludes ops labor, cooling, and redundancy — real self-hosting is worse than this, which makes a local win here a strong signal.

Does the volume pay for the price collapse?

The Jevons bridge between cheap intelligence and the capex bill. Total top-50 token volume per day, log scale — cheaper tokens are only bullish for everyone selling compute if demand more than makes up the difference.

Volume, 90d
+175%
6240B → 17162B tokens/day
Price, 90d
-68%
$0.45 → $0.15/Mtok (est.)
Implied spend, 90d
-11%
volume × price

Jevons breaking down: prices are falling faster than volume is growing, so total inference spend is shrinking — the revenue behind the capex bill is thinning.

Provisional: volume is measured across the whole window, but the price leg still rests on curated pre-launch estimates carried forward — measured price floors only start 2026-07-17, so a fully measured 90-day window opens around 2026-10-15. Until then the jevons-break turning point reads no value rather than trip on an estimate.

III

Durability

Does open stay open?

A cheap open majority only commoditizes intelligence if it keeps being downloadable. Three ways it stops without any change in the technology: the traffic concentrates in one jurisdiction, the weights get gated, or a government writes a rule. This is where the thesis is most likely to die.

Where the tokens actually run

Open-weight share of OpenRouter daily token volume — revealed preference, not benchmarks. Everything below is a claim about the durability of this share.

Whose open weights?

Chinese-published share of open-weight token traffic. Not a market-share stat — a concentration one: the portion of the open majority that a single government's export-control decision could withdraw.

Openness is a policy choice, not a property of a model. Beijing is reported to be weighing limits on foreign downloads of Chinese model weights while keeping API access open, and Washington is weighing restrictions on enterprise use of Chinese open models. Either would move this line without any change in the underlying technology — which is why the open-majority trip means less the higher this runs. Models are attributed to the jurisdiction of the publisher; community finetunes are usually built on a base model whose jurisdiction is what actually matters, so this is a floor on the true concentration, not a ceiling.

Can you still download them?

Open-weight status verified daily against the publishers' own repositories, weighted by traffic — not inferred from the creator's name. A model is only open if the weights are actually obtainable today, under a license that lets you use them.

Gated or withdrawn
0.0%
of open traffic — the trip condition
Announced, unverified
0.0%
claimed open, weights not yet public
Non-permissive license
6.7%
downloadable, but conditions apply
Unmapped
46.4%
no repo mapping yet — not assumed open
ModelWeightsLicenseRepository
GLM-5publicmitzai-org/GLM-5
DeepSeek-V4 Flashpublicmitdeepseek-ai/DeepSeek-V4-Flash
MiMo-V2.5publicmitXiaomiMiMo/MiMo-V2.5
Tencent HY3publicapache-2.0tencent/HY3
Nemotron-3 Ultra 550Bpublicother (conditions)nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
GLM-5.2publicmitzai-org/GLM-5.2
MiniMax-M3publicother (conditions)MiniMaxAI/MiniMax-M3

announced means claimed open-weight with no public repository resolving yet — a verification gap, not a restriction, and new releases sit here legitimately while weights are uploaded. Only gated and withdrawn (weights that resolved on an earlier run and no longer do) count toward the trip condition. Probed as of 2026-09-18.

Policy watch

Government action that would change what open-weight access means without changing the technology. Hand-curated from primary reporting; a reported intention is never counted as a rule.

  • 2026-07-21ReportedCNrestricts open weights
    MofCom consulting AI and chip firms on export controls for model weights
    China's Ministry of Commerce is reported to be consulting Alibaba, ByteDance and Zhipu on limiting overseas transfer of training data and on restricting foreign users' ability to download model weights, while continuing to allow overseas customers to access the models as hosted services. This is the precise inverse of the commoditization thesis: open-weight access becomes API access, and the open majority stops being withdrawable-proof. Source: Financial Times.
  • 2026-07-20ReportedUSrestricts open weights
    Administration weighing a ban on use of Chinese open-source models
    The administration is reported to be considering restricting US use of Chinese open-source models on national-security grounds, after lobbying from leading US AI firms. Would decouple US-measured open share from global open share — moatwatch measures global OpenRouter traffic, so a US-only restriction would show up here as a divergence between what is available and what is used. Source: Axios.
  • 2026-06-09In forceUSrestricts closed frontier
    Anthropic Fable and Mythos withdrawn on government security concerns
    Two frontier models were withdrawn from availability by the US government over security concerns. Recorded here because it is a measurement-integrity event, not only a policy one: the frontier gap is computed against the best available closed model, so removing a frontier model narrows the measured gap for reasons that have nothing to do with open weights catching up. Source: Reuters.
IV

The bill

Who pays for the buildout if intelligence is a commodity?

~$700B a year of hyperscaler capex is underwritten by scarcity that the first three parts say is disappearing. This part watches the two places that shows up: the fundamentals, and the price of everyone exposed to them.

The capex reckoning: spenders vs suppliers

Equal-weight hyperscaler basket (AMZN MSFT GOOGL META ORCL) vs semis (SMH), indexed to 100 two years ago. Rising together = the boom; falling together = the correlated unwind. Adjusted closes through 2026-09-17 — one session behind by construction, since Polygon's delayed end-of-day feed publishes a session's close overnight and the collector runs pre-open.

HyperscalersSemis (SMH)

Drawdown from trailing 90-day high

The unwind gauge: both baskets 15%+ underwater at once is the trip condition — a simultaneous repricing, not a rotation.

HyperscalersSemis (SMH)

Frontier silicon vs commodity silicon

Nvidia vs equal-weight AMD/AVGO/MRVL/MU/TSM, indexed. When open models commoditize inference, the market splits these two — this week's Marvell-down-Nvidia-flat pattern, as a series.

NvidiaCommodity silicon

The capex bill

Hand-curated from earnings guidance and analyst estimates, updated 2026-08-19. 2026 figures are estimates.

Top-5 capex 2026E
$802B
≈2× 2024, ~75% AI-related
Capex / operating cash flow
93%
up from 33% in 2023 — something gives at 100%
Semis EPS growth 2026E
+98%
consensus; the expectations cliff
Guidance cuts YoY
0
trip condition: any top-5 cut
2023
$141B
2024
$223B
2025
$403B
2026E
$802B
  • 2026-02: combined 2026 guidance crosses ~$700B, nearly double 2025 (CNBC)
  • 2026-06: capex-to-revenue gap widening; markets begin repricing debt-funded buildouts (Forbes)
  • 2026-07-16: SOX enters bear-market territory (-20% from late-June record) on Kimi K3 + Iran escalation
  • 2026-07-28: Alphabet raises 2026 capex to $195-205B at Q2; shares fall 7% and drag the other hyperscalers — the first time a capex RAISE was punished rather than rewarded (CNBC)
  • 2026-08: Q2 season closes with every top-5 name raising or reaffirming. Amazon to ~$220B, Alphabet $195-205B, Meta floor lifted to $130-145B on memory-price inflation. Big-5 2026 guidance now ~$775-800B, ~64% over 2025, ~75% AI-directed
  • 2026-08: Microsoft CY2026 capex reads $190B -> $175B, but this is a finance-to-operating LEASE RECLASSIFICATION, not a cut — CFO Amy Hood stated CY2026 investment expectations are unchanged. Recorded because a naive read of the headline would falsely trip capex-cut
  • 2026-08: Oracle closes FY2026 at $55.7B capex — 82.6% of its $67.4B revenue, against a $50B guide — and guides FY2027 to ~$95B gross / ~$70B net, funded by ~$40B of new debt and equity
  • 2026-08: semis + semi equipment post ~133% YoY Q2 earnings growth, ~44% of the entire S&P 500's earnings gain; sector 2026 revenue consensus $975B+ (+26%). SOX has rallied ~21% off its 2026-07-29 low, back out of bear-market territory

Method

How this is built, and what it got wrong.

The collector runs once a day, decides on its own, and discloses every judgment here rather than blocking on review. Estimates are labeled as estimates and can never trip a pre-registered condition.

Pipeline decisions

The collector decides autonomously and discloses here — classifications, data gaps, threshold changes, alerts.

  • 2026-09-18collectRun complete: 200 models, openShare=71.8
  • 2026-09-18collectJevons 90d (2026-06-20→2026-09-18): volume +175.0%, price -67.6%, implied spend -10.8% (window rests on curated price estimates) — Jevons breaking down: prices are falling faster than volume is growing, so total inference spend is shrinking — the revenue behind the capex bill is thinning.
  • 2026-09-18policyFrontier reference caveat — 2026-06-09 Anthropic Fable and Mythos withdrawn on government security concerns (Reuters). The measured gap compares the best AVAILABLE closed model; a withdrawn frontier model narrows it for reasons unrelated to open weights catching up.
  • 2026-09-18collectMarkets as of 2026-09-17: correlated-dd 1.0%, divergence 7.2pp, dispersion -2.4pp
  • 2026-09-18availabilityWeight availability: 7 models probed — 0.0% of open traffic gated/withdrawn, 0.0% announced but unverified, 6.7% under non-permissive licenses, 46.4% unmapped
  • 2026-09-18collectOpen-weight jurisdiction mix: 88.2% CN — cn 88.2%, us 10.1%, eu 0.0%, other 1.7%
  • 2026-09-18collectOpenRouter rankings 2026-09-17: open-weight share 71.8% of 18733.5B top-50 tokens
  • 2026-09-18classificationUnknown creator "nex-agi" in rankings — counted as closed-weights; add to classify.ts if wrong
  • 2026-09-18classificationUnknown creator "dots-studio" in rankings — counted as closed-weights; add to classify.ts if wrong
  • 2026-09-18classificationUnknown creator "stealth" in rankings — counted as closed-weights; add to classify.ts if wrong
  • 2026-09-18collectOpenRouter: 445 models
  • 2026-09-18collectArtificial Analysis: 200 models via https://artificialanalysis.ai/api/v2/language/models/free