When the chief executive of the most valuable chip company in the world left his job this week, the news arrived in the flattest container American law provides. Nvidia's Form 8-K, the document a public company must file when something material happens, says three things. Jensen Huang resigned as president, CEO and board member. Chief financial officer Colette Kress became interim CEO while keeping her old job. Harvey C. Jones became the first-ever board chair, and the board is searching for a permanent CEO.
What caught our attention was how many of the week's heaviest facts came in this form. Anthropic filed an amended registration for its IPO. Cerebras filed a quarterly report. For two years the price of an AI company was set mostly by things people told reporters: a round that leaked, a run rate quoted from memory. This week the biggest numbers came from documents that people sign and can be sued over. Shipping news pointed the same way, with the top-scoring model also the most expensive to run and Google reportedly scheduling price increases. Announcements were cheap this week. The disclosed numbers and the price lists were not.
Venture markets and private capital
The document everything else gets marked against
On November 10, Anthropic filed Amendment No. 2 to its S-1, the registration statement a company files before selling shares to the public. It offers 825,000,000 Class A shares at $70 to $78, with 801,351,354 from the company and 23,648,646 from selling stockholders. The company's IPO website says pricing is expected November 18 and trading November 19, subject to regulatory clearance. Anthropic is private until it prices, so we cover it here for what it does to marks on everything else.
Third-quarter revenue was $16,956 million, with an operating loss of $567 million (about negative 3%). The company puts run-rate revenue, a recent period multiplied up to a year, at about $81 billion as of October. Annualizing the third quarter alone gives $67.8 billion. Secondary summaries put the post-offering share count at 13,086,672,593, which implies a value of about $916 billion to $1.02 trillion, or 13.5 to 15 times annualized third-quarter revenue (11.3 to 12.6 times the October run rate). We have not read that share count off the cover page ourselves, so treat the valuation as an estimate.
Seven companies, including Google, Microsoft, Nvidia, TSMC, AMD, Broadcom and Micron, agreed to buy $26.0 billion of Class N stock at the IPO price, with no lock-up. A lock-up is a period after an IPO when insiders agree not to sell. The S-1/A restricts holders of about 15.5 billion shares for up to 180 days and co-founders and large holders for up to 365, with up to 3.6 billion shares eligible for early release. A secondary summary lays out release steps on days 27, 48, 62 and 76 after the start of trading. If trading begins November 19, those fall on December 16, January 6, January 20 and February 3. We read those steps secondhand, but the first falls inside 30 days and matters to anyone buying shares in a private AI company that will be compared to this one.
Where the large rounds went
Crunchbase's list of the largest US rounds for November 7 to 13 is overwhelmingly AI. Its week runs Friday to Thursday and overlaps our prior window, so we read it for direction.
The money sorts into three groups. Some feeds the labs; Mercor raised a $750 million Series D at $20 billion. Some builds plumbing for software agents; LangChain raised a $300 million Series C at $3 billion, and Rox raised $100 million for a sales agent. Some serves the physical world; Nominal raised $240 million at $3 billion. The largest round is Mirendil's $1 billion Series A for agentic AI in research and development, led by Kleiner Perkins and Clearlake. We found no product disclosure, so it is a bet on the team. No application-layer copilot without proprietary workflow data made the top ten.
Every round on the list is $100 million or more, so we cannot see the middle of the market. At the top, valuations run from $1.1 billion (Databento) to $20 billion. The outlier is Mavria, an AI infrastructure company founded by former Department of Energy staff, which launched with $130 million of pre-seed and seed money and now has a Paradigm-led round at a $1 billion valuation. Torus raised $500 million and launched a modular inference-compute product, Foundry, in the same announcement, which suggests infrastructure founders are raising against something specific to deliver. We found no new fund closes, no Y Combinator or a16z Speedrun demo day, and no private burn data inside the window. A DataHyena note linked by Innermost Loop says Nvidia joined about one in eleven US rounds of $100 million or more; we could not check the population behind that ratio.
Physical AI and robotics
Path Robotics raised a $300 million Series E at $2.45 billion for physical AI in manufacturing, starting with welding. Agile Robots of Munich announced a $500 million Series F on November 12 at $8.5 billion, led by SoftBank Group with Nvidia participating. Both back robots that do bounded industrial work, and neither prices a general-purpose humanoid. Capital here is going to work cells with metrics you can measure.
OpenAI and the listed comparables
The Wall Street Journal and Bloomberg report that OpenAI's third-quarter revenue was about $10 billion, up roughly 50% from the second quarter, with an adjusted operating loss of about $10.9 billion. The adjusted margin improved to negative 109% from negative 150%. These figures come from internal documents and not from a filing. Our November 8 issue reported a $1.25 trillion valuation on OpenAI's first $11 billion tranche. Against $40 billion of annualized revenue that is about 31 times, roughly double Anthropic's multiple at the top of its range.
Cerebras, listed in May and now a comparable, reported its third quarter on November 12 (8-K), resolving a date we flagged on November 8. Core revenue was $230.9 million and core cloud revenue $144.9 million, up 272% from a year earlier. It announced two new hyperscaler customers and 700 megawatts of data center capacity under contract. CNBC reports core growth slowed to 70% from 103% in the second quarter and that the stock has lost almost half its value since the IPO. CEO Andrew Feldman said the company expects to triple revenue in 2027. The 10-Q shows about $217.8 million of new non-cancelable data center lease payments signed in the fourth quarter.
CoreWeave reported a $132.5 billion backlog on November 10, and its stock rose 24.71% the next day (CNBC). No Federal Reserve decision fell inside the window; the next meeting ends December 9.
Singularity signposts
The most capable model carries the highest price per task
On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, one underlying model sold two ways. Fable is generally available, and Mythos goes only to verified organizations. Both list at $8 per million input tokens and $40 per million output, down from $10 and $50. A token is a chunk of text, roughly three-quarters of a word.
Artificial Analysis, an outside group that runs its own tests, ranks Fable 5.2 first of 73 or 74 models on its Intelligence Index with 65 at maximum effort, against 59 for Claude Opus 5.5 and 57 for GPT-6 Astra. It is also the most expensive model priced, at $26.54 per index task and about 47 minutes of work. SemiAnalysis said the model's habit of checking its own work raised cost per task by 31%. So price per token fell while the cost to finish a task rose.
Products that sell a finished result at a fixed price and route each step to the cheapest effort level that clears it get easier to build. Fable 5.2 at medium effort scores 56 at $3.80 a task, above Fable 5.1 at maximum effort (49 at $12.99). Per-seat products that quietly pass maximum-effort usage through get harder to defend, because the customer controls the dial that sets their margin. We will watch whether Artificial Analysis or Epoch AI reproduces the 31% figure.
A robot assembled a structure in orbit
Caltech reports that its ASTRA demonstration is the first robot to assemble a structure in orbit autonomously, a truss bay of the type future large telescopes need (via Innermost Loop, November 10). It is a demonstration and not a service. If it generalizes, building in space no longer needs an astronaut or a ground operator in the loop. Startups selling autonomy software, verification tools and standard truss or docking hardware benefit, which fits our deeptech interests. Plans that assume large structures must launch fully assembled look weaker. Look next for a repeat with different geometry or a NASA or Space Force solicitation naming autonomous assembly.
A general robot model hits a factory-grade number, by its maker's account
Google DeepMind says Gemini Robotics 2.5 sorts Mercedes parts with a 96% success rate. No outside party has reproduced it; a published throughput and intervention rate from Mercedes over a full shift would settle it. If it holds, a general model has reached the reliability at which a line manager will pilot it without custom training. Integrators that wrap a general model around one cell, such as kitting or inspection, and keep the deployment data gain. Startups whose edge is a proprietary model for that same cell lose ground.
We left out forecasts, such as the AI Futures Project moving its intelligence-explosion estimate to 2027. They change no capability, cost or deployment fact inside the window.
Foundation and open-source model watch
Moonshot AI's Kimi K3.1 Pro weights are out, resolving the November 4 date our November 1 issue flagged. The license is custom, not Apache 2.0 or MIT, so commercial use needs diligence. The model has 2.8 trillion total parameters with 104 billion active (a mixture-of-experts design, where only part of the network runs for each token) and a 1-million-token context. Any seed pitch whose edge is frontier-scale agentic capability available only through closed APIs got thinner.
The bigger story for Chinese open labs was chips. Bloomberg reported on November 10 that Moonshot rented Nvidia's top B300 chips inside China at a 350% markup, and The Information reports Commerce is finalizing a rule barring Chinese AI firms from renting advanced chips in third-country data centers. If it lands, training the next DeepSeek, Qwen, Kimi or MiniMax release gets more expensive. Weights already released stay free, but the pressure Chinese models put on Western mid-tier pricing could ease within two or three release cycles.
Meta says its Muse Ember matches Claude Haiku 5.5's best score at a third of the cost. The post names no benchmark, so an Artificial Analysis entry with cost per task and a stated license would settle it. We found no primary releases this week from DeepSeek, Qwen, MiniMax, Mistral, Z.ai, Xiaomi MiMo, Ant Group, Nvidia Nemotron, Reflection or Arcee.
Platform power and incumbent moves
ABC News reports Huang's resignation followed days of reporting about an alleged affair with a woman US officials suspect of working for Chinese intelligence, and Reuters reported a relationship with a second employee. These are allegations in press reports and not findings. CNBC puts the stock about 30% below its peak, and earnings are due next week. Elon Musk posted that hiring Huang would be an honor, so the destination, if there is one, is open. A founder-CEO exit at a top-five company is rare enough that we treat it as one event and not a trend. Alternative silicon and inference clouds that pitch supplier diversity had a better week. Startups whose distribution depends on Nvidia's partner programs and venture arm now rely on relationships the person who built them no longer oversees.
Anthropic launched Claude Meetings on November 10, which joins workplace meetings and turns the discussion into tasks, and added Sheets and Whiteboards (TechCrunch). Standalone meeting-notes tools and lightweight AI spreadsheet add-ons have a harder road. Vertical systems of record, which a general assistant must connect to in order to act, have an easier one.
Google launched Gemini 4 Hydrogen on November 12 across its API, app and Search, where it is the new default in AI Mode. A default for every Search user squeezes consumer answer engines unless they own a vertical source of data.
Reuters reported on November 9 that a judge ruled Anthropic must face a class action over how it marketed "5x or 20x more usage" on Max plans. Any startup that resells model access with multiplier language should look at its copy.
Compute and inference economics
Anthropic's IPO factsheet discloses about $900 billion of compute commitments as of and after September 30. About $451.6 billion is expected before the end of 2030, and about $70.7 billion comes due in the next 12 months. Google and Amazon deals are non-cancelable except for uncured material breach, and Microsoft's is the same with a carve-out for certain undelivered tranches.
The SpaceX arrangement has unusual terms. Anthropic agreed to spend up to $120.5 billion into 2029, about $3.1 billion a month once all capacity is delivered. After non-cancelable periods ending in January and March 2027, either side can cancel on 90 days' notice (filing). After March, SpaceX holds the exposure on that capacity if Anthropic leaves, and our November 1 issue reported SpaceX's own roughly $40 billion GPU lease runs about three years past the customer contract. The loss turns real if Anthropic gives notice and SpaceX cannot re-rent the capacity at a similar price. A report citing The Information says Anthropic does not expect the arrangement to last beyond the second quarter of 2027; the filing is silent, so we treat that as unconfirmed.
We owe readers a correction. Our October 25 issue said the commitments equaled about fourteen years of revenue, pairing an outdated revenue figure with a commitment number that appears to match only the through-2030 slice. The two errors roughly cancel, which is why the right answer lands nearby. The full $900 billion is 13.3 years of annualized third-quarter revenue and 11.1 years of the October run rate. The slice due by 2030 is about $106 billion a year, 1.3 times the October run rate. With a third-quarter operating loss of only $567 million, current revenue covers most current compute, and the risk sits in the growth assumptions behind later years.
Fable 5.2 and Hydrogen changed the cost-per-task ladder, so here it is, sorted by cost on Artificial Analysis's Intelligence Index v5.0.1, with figures compiled by outside trackers.
Model and effort
Index score
Cost per index task
Claude Fable 5.2, medium
56
$3.80
Claude Fable 5.2, high (API default)
62
$9.08
Claude Opus 5.5, max
59
$10.04
Claude Fable 5.1, max
49
$12.99
Claude Fable 5.2, max
65
$26.54
Fable 5.2 at high effort beats Opus 5.5 at max on both score and cost, and the last three index points cost almost three times as much. Gemini 4 Hydrogen scored 54 but has no cost-per-task figure yet.
Hydrogen lists at $1.50 per million input tokens and $7.50 per million output, replacing Gemini 3.8 Flash, with Google claiming 2.5 times Flash's speed. Trackers reading Google's pricing page report that Gemini 4 Carbon and Argon go from $2 and $10 to $4 and $20 on December 1, and that Flash rises to Hydrogen's price on January 1. If Google's page confirms it, this is the first scheduled frontier price increase from a major lab we have logged, and it breaks the assumption that model input costs only fall.
On silicon, Epoch AI finds a GB300 delivers about 25 times an H100's agentic inference throughput on 2.5 times the raw arithmetic, because agent workloads are limited by memory bandwidth. A product billed per token on an older fleet can lose money on long agent sessions that a newer fleet serves at a profit, so startups that rent inference should ask which generation their provider runs. OpenAI's negative 109% margin means about $2.09 of adjusted operating cost per revenue dollar, against Anthropic's negative 3%, which hints at which supplier faces more pressure to raise prices or ration capacity.
Macro, regulation and physical infrastructure
The Bureau of Labor Statistics reported on November 10 that October CPI rose 0.4% after 0.6% in September, with core at 0.3% against a consensus of 0.2% (CNBC). BLS published no 12-month changes because the October 2025 lapse in appropriations prevented data collection, so we cannot extend the 3.7% September figure we ran on October 18. The next CPI release, December 10, comes a day after the December 8 to 9 Fed meeting ends, so the Fed decides without November data. A hot core print makes a December cut less likely and keeps discount rates where they are for private equity.
Senate leaders see a "high probability" of tighter export controls after the Nvidia scandal, per Semafor, which adds to the pending Commerce rule. Startups building know-your-customer and usage-attestation tools for GPU clouds gain a regulatory buyer if it lands. Politico reported that New York Governor Kathy Hochul launched the first state AI testing program. The Nuclear Regulatory Commission issued a permit for X-energy's Xe-100 plant in Texas, the first for a high-temperature gas-cooled reactor in nearly 60 years. Bloomberg reports SpaceX plans a $24 billion data center and gas plant at Starbase, in a state whose governor-elect campaigned on curbing data centers. The Pentagon is reportedly lending launch support to SpaceX rivals as Falcon winds down (Ars Technica), widening demand for dual-use launch and in-space assembly suppliers.
Cross-stack interaction effects
Anthropic's compute bill and Cerebras's order book
Anthropic's filing shows a SpaceX arrangement that becomes cancellable in early 2027. In the same week Cerebras reported two new hyperscaler customers and 700 megawatts under contract. The largest buyer's own filing now tells every supplier when part of its spending can move, which gives alternative-silicon vendors a date to pitch against. Inference clouds on non-Nvidia chips and workload-portability tools gain. GPU lessors whose lease terms run past their anchor tenant's cancellation date lose. Cerebras trades near half its IPO level while guiding to triple revenue in 2027, and we see no sign anyone has priced that window as a sales opportunity. Horizon is three to nine months.
Nvidia's leadership change and the Commerce rule
A newly led Nvidia has every reason to over-comply with a rule against third-country chip rentals, and the rule gives it the legal footing. Demand that reaches China through intermediaries would shrink faster than either change implies alone. Compliance tooling, American open-weight labs (Reflection, Arcee, Nemotron) and non-Nvidia accelerators gain. Neoclouds with Asian reseller exposure lose, as do startups whose cost model assumes Chinese open weights keep improving monthly. The market is pricing the person and has not priced what this does to open-weight competition. Horizon is six to eighteen months.
Google's price increase and Fable 5.2's cost per task
If the December 1 increases hold, price per token at the frontier and tokens per task are rising together, a first in anything we have logged. Routing products, evaluation tools that prove a cheaper tier is good enough, and fixed-price outcome businesses gain. Seat-priced products with uncapped frontier usage lose. This starts December 1.
What this means for founders
More attractive now
Effort and routing control, since a product that picks the model and effort for each step can deliver index-56 work at $3.80 instead of index-65 work at $26.54. Inference on non-Nvidia silicon and the tooling to move workloads between chips. Workflow software for one robot cell that sits on a general robot model and keeps the deployment data. Compliance and attestation for compute access, which gets a buyer the day a rule is published. Autonomy and verification software for in-space assembly.
Less attractive now
Standalone meeting-notes tools, consumer answer engines without proprietary data, seat-priced AI tools that pass frontier usage through without caps, and GPU leasing whose terms outlast the anchor customer's cancellable period.
Worth watching, with caution
Intelligence-explosion forecasts change nothing in a contract or price yet; watch whether any lab's spending guidance starts to cite them. Agent-led science as a consumer pastime, such as the Claudestronomy dwarf-planet candidate, produces real finds and no visible business model.
Worth more attention than it gets
The calendar of Anthropic's cancellable compute tells every competing supplier when to pitch. Artificial Analysis's v5.0.1 scores are not comparable with v4.x numbers, so any deck quoting an index score should state the version. With no November CPI before the December meeting, rate-sensitive startups should plan for no cut. And every 2027 model-cost line built on "prices fall 50% a year" needs a downside case.
Questions for founders this week
- If Gemini 4 Carbon doubles on December 1 and your top model's cost per task rises 31%, what is your January gross margin, and which steps can drop to a cheaper tier without a measured quality loss?
- Which hardware generation serves your inference, and what happens to long-session agent margins on older chips?
- If you sell into factories, what is your intervention rate per shift, and would you beat a model claiming 96% on a sorting task?
- If the remote-access rule passes, which of your providers, resellers or customers would fail a know-your-customer check?
- If Anthropic trades at the top of its range, does your next round's comparable set change?
Secondary-market watch list
- OpenAI: a $1.25 trillion private mark beside a filed public peer at roughly half its revenue multiple.
- Mercor: $20 billion on a $750 million Series D; exposed to lab spending cycles.
- LangChain: $3 billion on a $300 million Series C; agent governance as frontier tokens get costlier.
- Path Robotics: $2.45 billion on a $300 million Series E; a narrow industrial physical-AI comparable.
- Agile Robots: $8.5 billion on a $500 million Series F led by SoftBank.
- Nominal: $3 billion on $240 million; engineering tools for defense and physical-AI supply chains.
What this means for Team Ignite LPs
The seed rounds that matter to us are routing and effort control, compute compliance, robot-cell workflow software and alternative-silicon tooling.
The late-stage AI names we follow are about to get a public reference when Anthropic trades. Any secondary priced off the IPO range over the next 90 days should assume more shares reach the market at each release step, and the $26 billion placement has no lock-up at all. Physical AI rounds this week priced narrow, measurable industrial work, which supports tilting robotics exposure toward cells with clear throughput numbers. FDA-gated rounds such as Delix's $100 million Series C stay outside our thesis however large they are. The message to LPs this quarter is that we are moving from narrative marks to filed comparables, and some marks may reset in either direction once Anthropic trades.
What this means for VCs
Three things to watch. Anthropic's first weeks of trading, because every private AI lab multiple will be read against it. Nvidia's earnings call under interim leadership, including whether its venture arm keeps its one-in-eleven pace in large rounds. And Google's December 1 price change, which will show whether frontier increases stick.
Funds concentrated in Nvidia-adjacent strategic co-investments should map how many portfolio companies depend on Nvidia partner programs or capital. Funds heavy in per-seat application pricing should model margin under rising per-token prices and rising tokens per task at once. Two places look mispriced. OpenAI private paper sits against the Anthropic range, and billion-dollar seed and Series A rounds for unproven teams, such as Mirendil and Mavria, are priced as if forecasts of imminent takeoff were contracts.
