Public markets began pricing AI companies by how they pay for compute this week, and three model launches cut the price of renting that compute in the same seven days. Alphabet grew revenue 25% to $127.6B and still had its worst trading day on record after reporting negative free cash flow. SpaceX committed to a roughly $40B GPU lease with an Apollo-managed entity. Anthropic's public S-1 named its compute counterparties and left the total blank.
Over the same days, GPT-6.1 Terra, Gemini 4 Carbon and the promised Kimi K3.1 Pro open weights pushed the price of a benchmark-grade task down. Intelligence got cheaper to rent and dearer to finance. Companies that own the compute carry the financing risk. Companies that rent it collect the deflation. Seed-stage application companies sit on the second side.
Three takeaways for us:
- Favor seed companies whose gross margin rises when model prices fall, and avoid those that resell mid-tier capability at a markup.
- Physical AI is paying for proof of unattended work, and the October 28 tariff proclamation on robotics favors software that upgrades machines already in the field.
- For secondaries, wait for the Anthropic S-1/A price range, and treat GPU-lessor exposure as credit risk with a re-contracting date, not as equity upside.
Venture markets and private capital
Physical AI and exchange infrastructure drew the largest checks. Every round on Crunchbase's October 24 to 30 list is nine or ten figures, so we have leader data only and cannot say how the middle is pricing.
Bedrock Robotics raised a $600M Series C at $4.1B post-money, co-led by CapitalG and Emergence Capital. The February Series B was $270M at $1.75B post, so the price rose about 2.3x in eight months. The trigger was a milestone, not a demo: its excavators have done paid work with no operator in the cab since mid-August. In robotics, the event that unlocks the next price is a customer paying for work nobody supervised.
Kalshi raised a $1B Series G at $40B post-money, co-led by Wellington and Sequoia. The company says annualized volume exceeds $1 trillion. That figure is company-reported and unaudited.
Also on the list: Re:Build at $2.1B, co-led by General Catalyst and JPMorganChase, and Serval at $150M on a $3B valuation (coding agents for enterprise automation).
Listed comparables moved against holders:
- Alphabet's Q3 free cash flow was negative $9.07B against 2026 capex guidance of $195B to $205B. Search grew 12%, down from 17% in Q2, while Cloud grew 103% (Alphabet release). The stock closed down 12.6% at $283.82 on October 29, passing the 11.63% drop of March 2020 (CNBC).
- The same CNBC report has the Global X AI & Technology ETF down 8% in October and the iShares Semiconductor ETF 29% below its recent high.
- SpaceX's Q3 filing shows a "Customer B" at 34.5% of revenue, which Bloomberg reports is Anthropic. Neither company has confirmed it.
On rates, the FOMC held the federal funds range at 3.75% to 4.00% by a 10 to 2 vote on October 28. Governor Lisa Cook and Cleveland Fed President Beth Hammack dissented in favor of a 25 basis point hike. There was no projections release at this meeting. Fox Business cited CME FedWatch at 82.6% for a December hike on October 28, and the reading had eased to 70.7% by October 30. The next meeting is December 8 to 9. A higher discount rate and lower public comparables both cut late-stage private marks this week.
We found no fund closes at our stages and no YC or a16z Speedrun pricing data dated inside the window.
Singularity signposts
The benchmark got harder and the top spot changed hands in a day
Artificial Analysis released Intelligence Index v5 on October 27 with two new benchmarks, three upgraded ones, and most of the index now built on private datasets. Claude Opus 5.5 led at 58. On October 28, Google said Gemini 4 Carbon ranks first at 59, ahead of Opus 5.5 at 58 and GPT-6 Astra at 57. Carbon is available to cyber defenders only for now. Opus 5.5 still leads the Coding Agent Index, 62 to 59.
The ranking holds across all models on v5 as of October 28, and a one-point lead on a composite is inside the range where effort settings can flip the order. The rebuild matters more than the winner. Labs can no longer climb an index that tests on private data by tuning to public test sets, and scores from different index versions cannot be compared. A seed deck that cites a v4 score may overstate a model by several points on v5.
This makes evaluation products that score a customer's own workload more defensible, and it weakens any pitch that quotes an index number without naming the version. Watch for Carbon's general availability and price.
Mid-tier intelligence got a new price floor
OpenAI released GPT-6.1 Terra on October 26 at $1 per million input tokens and $5 per million output, half the per-token price of GPT-5.6 Terra and GPT-6.1 Sol. OpenAI says Terra matches GPT-6 Astra on its computer-use evaluation at one-tenth the per-token price.
Per-task cost is the more useful number. On Artificial Analysis v5, Terra at max effort scores about 51 at roughly $0.96 per task, against about 54.5 and $1.27 for GPT-6.1 Sol (see the table in the compute section). Back-office agents for invoice entry, document routing and screen-reading QA now fit under a dollar per task, which supports SMB pricing at healthy margins. Startups that earned margin by reselling mid-tier capability at a markup lose it. Watch whether Anthropic or Google reprices its mid-tier within 30 days.
Open-source models and platform moves
Kimi K3.1 Pro weights are due November 4
Kimi K3.1 Pro launched October 23, and we reported its $1.98 cost per task and top-four rank on October 25. The new information is the weights date. Moonshot says open weights arrive by November 4. On the prior index version, Artificial Analysis says that would narrow the gap between the best open and proprietary models from 12 points to 3. It measured K3.1 Pro at 67% on Terminal-Bench 4.0, the highest of any model it tested. The model has 2.8T total parameters, 104B active, a 1M context window, and API pricing of $2 input and $10 output per million tokens.
The license is the open question. K3 shipped under a custom license rather than an OSI-approved one, and K3.1 has no announced terms. If the weights ship under usable terms, two moats get thinner: self-hosting and data residency as a closed-lab selling point in regulated mid-market deals, and the coding-agent advantage of tools that depend on one closed frontier model. A company whose pitch is hosting or fine-tuning a frontier-class model for customers who cannot send data out would then compete with a free download. The ranking is open-weights-only, on the old index. Opus 5.5, Sonnet 5.5 and Muse Wonder Full all scored higher on v4. Artificial Analysis's model page now shows 46 for K3.1 Pro, which we believe reflects the v5 rebuild.
We checked only Artificial Analysis as a discovery channel this week and found no verified releases from the other open-model labs.
Platforms
- GitHub Copilot Fusion routes between Anthropic and OpenAI models and, per Artificial Analysis, scores 2 points below Opus 5.5 in Claude Code at half the cost per task. Routing now ships inside the default IDE assistant, which squeezes standalone coding-model routers and leaves room for evaluation and governance tools that sit above any orchestrator.
- Microsoft's 10-Q shows $10B of revenue from arrangements with OpenAI in the quarter ended September 30, about 11% of $92.3B, with OpenAI receivables of $7.2B.
- Alphabet's Search growth of 12% came in a quarter when query growth was at a five-year high. Revenue per query is falling as AI answers absorb clicks. That helps agentic-commerce and answer-engine tools and hurts performance-marketing startups built on search arbitrage.
Compute and inference economics
What Anthropic's S-1 adds
Anthropic's S-1, filed October 26, lists Nasdaq as the venue, "ANT" as the proposed ticker, and Goldman Sachs and Morgan Stanley leading a 24-bank syndicate. Price range, share count and Q3 revenue are blank. The filing reports:
- Net revenue retention of about 900% as of September 30 (excluding Pro and Max subscribers).
- More than 11,900 customers above $100K of trailing revenue, more than 1,900 above $1M, and about 250 above $10M.
- Second-quarter revenue growth of 1,368% year over year against 300% growth in compute and infrastructure expense.
- Commitments running through 2045, with substantially all of the remainder incurred between 2031 and 2045. The aggregate is left blank.
Governance matters for anyone pricing a secondary. One Class F share held by a founder LLC carries 50.1% of general voting power, the Long-Term Benefit Trust elects a board majority through Class T stock, and strategic partners hold non-voting Class N stock. Buyers of Class A get neither voting control nor board control, and a price should reflect that.
Commitments, by structure
The commitments share a headline and differ in who bears the risk. Amounts below come from secondary readings of the S-1 and should be checked against the filing.
Counterparty
Amount
Term
Structure
about $161.2B
five years
Equipment lease commitment, noncancelable except on default
$114.5B
2026 to 2033
Cloud purchase commitment, payable regardless of usage
$110B
2026 to 2036
Cloud purchase commitment, payable regardless of usage
up to $120.5B
into 2029
About $3.1B a month once capacity is fully delivered in December
$44.9B
2026 to 2033
Cloud purchase commitment
The named arrangements sum to about $551B. An additional TPU procurement of $125B to $150B over 2028 to 2031 takes the total to roughly $676B to $701B. We published $518B on October 4 and about $450B on October 25. The October 25 number counted cloud commitments only, so the gap is definitional.
The arithmetic needs the latest revenue figure. Q3 revenue of $16.8B, which we reported on October 18, annualizes to $67.2B, and $551B is about 8.2 years of that. Our October 25 estimate of about fourteen years used first-half revenue of $16.3B and overstated the ratio. That was our error.
SpaceX's side of the Anthropic contract also grew. The version we covered earlier ran about $45B with a 90-day termination right (Teslarati). The S-1 says the agreements were amended between May and October and now total up to $120.5B. We have not seen whether the termination right survived, and it is the most important open question for SpaceX's AI segment.
Who holds the exposure on SpaceX's GPU lease
SpaceX's 8-K describes a binding commitment for a roughly $40B master finance lease for GPUs, including Nvidia's Vera Rubin platform. Each tranche runs about five to five and a quarter years from delivery, delivery starts in the first half of 2027, and closing is expected before year-end. Musk posted on October 25 that the debt had closed, while Bloomberg reported pricing talks were still ongoing. The filing's language is "binding commitment."
- Tenor: the lease runs to roughly 2032. The customer contract that justifies the capacity runs into 2029.
- Exposure: SpaceX owes the payments. The Apollo vehicle holds GPU residual value plus SpaceX credit risk.
- Loss condition: if Customer B does not renew past 2029 and replacement demand prices below the lease cost, SpaceX equity absorbs the gap for about three years. The lessor loses only if SpaceX cannot pay.
Musk's claim that debt markets have almost shut for companies without non-data-center revenue is an interested party's assertion. We have no lender-side data on it.
Cost per task on index v5
Benchmark-adjusted cost sorted low to high.
Model (effort)
v5 score
Cost per task
GPT-6.1 Terra (medium)
38.1
$0.17
GPT-6.1 Terra (max)
51.0
$0.96
GPT-6.1 Sol (max)
54.5
$1.27
Gemini 4 Carbon (high)
59
$7.76
Claude Opus 5.5 (max)
58
$10.04
About 51 points costs under a dollar, and the last 7 or 8 points cost 8 to 10 times more. Most production agents should route to the mid-tier by default.
Macro, regulation and physical infrastructure
On October 28 the President signed a Section 232 proclamation finding that imports of robotics, industrial machinery and their parts threaten national security. It directs Commerce and USTR to negotiate with trading partners and report progress within 180 days, which is April 26, 2027. No tariffs or quotas take effect now. KPMG notes that tariffs, duties, quotas, fees and a domestic-production incentive program remain available as later remedies. Contract manufacturers will likely start asking robotics vendors for country-of-origin detail in 2027 procurement. The rate decision is covered under venture markets above.
We did not check state utility commissions or data center siting actions this week.
Cross-stack interaction effects
Alphabet priced the cost of AI answers, and OpenAI cut the price of the agents producing them
Alphabet's Search grew 12% in a quarter when Cloud grew 103%, with query growth at a five-year high. In the same week, GPT-6.1 Terra arrived at $1 and $5 per million tokens, pitched at computer-use and browsing tasks. The cheapest capable agent shipped as the largest ad-funded search business showed that agentic traffic does not pay like human clicks.
Agentic commerce, merchant-side agent readiness (structured catalogs, agent-facing checkout) and answer-engine analytics benefit. SEO and paid-search arbitrage businesses, and content sites that live on Google referrals, lose. Public markets repriced Alphabet in one day. Private marks on ad-tech and SEO tooling have not moved. Horizon: 6 to 18 months.
A five-year GPU lease sits on a customer contract that ends in 2029
SpaceX's Apollo lease runs about five years from 2027 deliveries. At the same time, Terra and the pending Kimi weights pushed task costs down, and Anthropic's S-1 claims Opus 5.5 cut customer cost per task by about 40% versus Opus 5 (a vendor claim). The market rate for the compute behind a lease is set by model efficiency curves, and those fell again this week.
Inference-optimization tooling and senior lenders benefit, along with hyperscalers that can absorb underused capacity from operating cash flow. Single-customer neoclouds and anyone marking SpaceX's AI segment as recurring revenue are exposed. SpaceX equity trades on segment growth while the 2029 to 2032 re-contracting risk sits in a footnote. Horizon: 2 to 5 years.
The tariff clock and the retrofit round point the same way
The Section 232 proclamation starts a 180-day negotiation on imported robot arms and machine tools. Bedrock's $600M round funds autonomy kits for machines already working on US sites. If tariffs follow in 2027, new imported hardware costs more and software that upgrades the installed base gains value.
Retrofit autonomy, US-assembled robotics, and fleet and maintenance software benefit. Robotics startups whose bill of materials depends on imported arms and reducers are exposed. Seed rounds in hardware-heavy robotics do not yet price that exposure. Horizon: 6 to 24 months.
What this means for founders
More attractive now
- Back-office agents priced per completed task on mid-tier models, where a sub-$1 task cost supports SMB pricing at strong gross margin.
- Retrofit autonomy for heavy equipment, agriculture and industrial machines already in the field.
- Agent-readiness tooling for merchants and publishers: structured catalogs, agent-facing checkout, attribution for agent-originated purchases.
- Private, workload-specific evaluation services, now that the leading public index moved to private datasets.
- Self-hosting and compliance tooling for open-weights frontier models, if Kimi K3.1 Pro ships on commercial terms.
Less attractive now
- Standalone coding-model routers, with Copilot shipping orchestration at half the cost per task.
- SEO and paid-search arbitrage.
- Single-customer GPU resellers and small neoclouds that need debt.
- Hosting a frontier model for regulated customers as a sole moat, if weights ship November 4.
- Robotics hardware built on imported components with no domestic assembly plan.
Overhyped but worth watching
- AI-run legal entities. A Delaware bill is drafted, and the legal path is at least a legislative session away.
- One-point leads on composite indexes. Carbon's 59 against Opus's 58 is inside run-to-run noise, and Carbon is gated.
- Prediction-market valuations, which rest on company-reported volume.
Under-discussed
- Tenor mismatch in AI infrastructure: five-year equipment leases on top of customer contracts that end sooner.
- Revenue per query, as opposed to query volume, as the measure of AI substitution in ad businesses.
- Index version drift in decks.
Questions for founders this week
- If your main model's per-task price fell another 50% next quarter, does your gross margin rise, or do customers ask for a price cut?
- Which of your benchmark claims name an index version, and do they hold on v5?
- If Kimi K3.1 Pro weights ship November 4 on permissive terms, what would a customer self-host instead of buying from you?
- For hardware founders: what share of your bill of materials is imported robotics or machine-tool content?
- For anyone selling to advertisers: what happens to your pipeline if Google's search growth stays near 12%?
Secondary-market watch list
- Anthropic: S-1 is public. Watch the S-1/A price range and the discount for Class A's lack of control.
- Bedrock Robotics: $4.1B post-money, the physical-AI benchmark for retrofit autonomy.
- Kalshi: $40B post-money. Regulatory exposure is the swing factor.
- Serval: $3B valuation in enterprise coding agents, a category Copilot Fusion now pressures.
- Waymo: Alphabet's results raise questions about parent capital allocation. Watch for any change in external funding appetite.
What this means for Team Ignite LPs
Our stage sat on the beneficiary side this week. Our companies rent intelligence and own no compute, so each cut in task cost, like Terra's 50% price drop, flows into their gross margin, and they carry none of the financing risk that public markets just repriced. The October 26 to 30 selloff was a capex-financing selloff, and seed-stage application companies are buyers of the deflation that capex produces.
The secondary book faces two inputs that both moved against holders. The Fed held with two hawkish dissents and a market-implied December hike, and the most liquid AI comparables re-rated, with Alphabet's worst day on record and semiconductor ETFs down 29% from their high. Any secondary priced off pre-October comparables should be re-underwritten before closing. For Anthropic, the S-1 replaces rumor with a filing, and the right move is to wait for the S-1/A price range and to price in the Class F and Long-Term Benefit Trust structure.
For portfolio construction, physical AI deserves a deliberate allocation going forward. The market paid for Bedrock after it proved unattended commercial work, and small initial checks can enter before that milestone, provided we diligence each bill of materials for Section 232 exposure.
For you as LPs, the message is this. Public AI companies are being sorted by whether operating cash flow covers their capex, and our stage sits on the beneficiary side of that sort. Late-stage secondary marks currently face a higher discount rate and lower public comparables.
What this means for VCs
Five dates to watch: the Anthropic S-1/A price range and first-day float, the Kimi K3.1 Pro weights on November 4, Gemini 4 Carbon general availability and pricing, the December 8 to 9 FOMC, and Alphabet's 2027 capex guidance on its Q4 call.
Funds with large growth-stage AI infrastructure exposure carry financing-structure risk their marks may not show. Funds concentrated at pre-seed and seed in application-layer AI hold the hedge. Crossover investors who just sold Alphabet on negative free cash flow will likely apply the same screen to private rounds, so expect tranched closes and milestone pricing above $1B.
We see three places where pricing looks off:
- Physical AI at seed. Late-stage prices have run ahead of seed prices for companies one milestone behind (Bedrock rose about 2.3x in eight months), and the tariff tailwind is not yet in seed pricing.
- Private ad-tech and SEO tooling marks, which have not absorbed what Alphabet's Search growth disclosed.
- GPU lessor exposure marked as growth equity. In practice it is credit risk with a re-contracting date around 2029.
Coverage: October 26 to November 1, 2026.