How DeepStack settles a question · Part 12
Does the frontier model still command a price?
From Part 12: The Model Is Not the Price. 21 voices on 3 sides, attributed and dated; the DeepStack Call with its history; 6 dated tests that settle it; the results as they come in.
The mechanism
The token is cheap and getting cheaper. The gap between the Chinese open models and the frontier is about four months by Torsten Slok's reading of Epoch, a majority of tokens on the platforms that count them had moved to open or Chinese models by September, and Anthropic's Opus 5.5 lists at $4 per million input tokens against OpenAI's $10, a 60 percent undercut. What still commands a price is the lab: the completed job, the distribution, the dollars that have not yet followed the volume. But no audited figure for that price is on EDGAR; $50 billion, $65 billion and $70 billion are three run-rate readings on three dates, and only a registration statement puts them on one line. The Call expects the filing to show a lab that is paid, in part by its investors.
Approved by Marco Cauduro on October 10, 2026.
The DeepStack callSplit
A frontier model no longer commands a price per token for long. The lab still does, and only a prospectus says what that price is worth.
Our Call takes the side “The filing sets the price”.
Call history
- v2The price has left the modelConviction Moderate (3/5)Side: The filing sets the priceSide named by the owner.
- v1The price has left the modelConviction Moderate (3/5)First call, with the story.
11 voices
The frontier is still paid
Reported revenue acceleration at the two leading model companies — an annualised run-rate said to go from USD 1 billion at end-2024 to USD 9 billion at end-2025 to almost USD 50 billion in May this year at Anthropic, with a similar trajectory at…
Tae Kim anticipates a close back-and-forth for model leadership between OpenAI and Anthropic over the next six months, expects another Anthropic model release before or around its IPO, and predicts both American frontier labs should do great once…
Tae Kim argues the earlier media consensus that OpenAI had fallen too far behind Anthropic was wrong, and that OpenAI pivoted resources to agentic coding after Anthropic's December lead, largely caught up by April, and arguably surpassed Anthropic…
Oguz Erkan says the concentration of the LLM market is still unknown: OpenAI and Anthropic take the lion's share of token dollars but their share of token volume has been declining, so the LLM market could be far more fragmented in 3-5 years, which…
Wyndo believes an always-on agent should run on the most capable model available, which is why they consider it fitting that ChatGPT's Dots runs on GPT-6 Astra.
Wyndo recommends trying open-weight models if a user keeps hitting usage limits on Claude or ChatGPT, or wants to lower the cost of certain tasks, since some work may run well on an open-weight model while saving the subscription for tasks where…
What actually fell is cost per completed job rather than cost per token: against Sol, Astra finishes Terminal-Bench 4.0 tasks at roughly 9% lower estimated API cost per task and OSWorld 2.0 tasks in about 47% less time, so the relevant demand…
They judge the switching-cost lever basically unavailable because the AI frontier is a Red Queen's Race at the model layer, with swapping between models easy down to task-by-task switching and everyone improving continuously.
They say a pretty small group uses AI rampantly for everything and a somewhat larger group uses ChatGPT regularly, meaning most people have not yet been awakened as customers, and they think OpenAI understands how to do it.
Criteo launched an Agentic Commerce Recommendation Service letting AI systems use Criteo's commerce intelligence to rank products, and OpenAI selected Criteo as its first advertising technology partner for ChatGPT Ads, with more than 2,000 brands…
Based on their capex sustainability model, Market Sentiment estimates that next year's ~$1.2T+ of AI capex should drive ~$2T of incremental revenue, which they say will only be a fraction of the total incremental revenue AI generates as a business…
10 voices
Open weights set the price
Chinese AI models captured a majority share of 57% to 67% of total token usage on the OpenRouter developer platform by September 2026, eroding OpenAI's token marketshare even as OpenAI subsidizes its new cheaper GPT-6.1 Sol model.
Anthropic is far ahead of OpenAI pre-IPO, both in model quality and ARR, and that lead should only accelerate with IPO proceeds heading into its November IPO.
OpenAI's price cut on its GPT model, combined with a reported tenfold jump in usage, looks like the beginning of a price war in AI services.
Steve says the AI story probably continues as long as token growth keeps rising, and identifies a price war between Anthropic, OpenAI and the Chinese models as the main risk that could change it.
China is gaining ground in AI, with Chinese models now accounting for 66% of token share.
Under 2% of tokens to a majority.
China is releasing open frontier models not despite state involvement, but because an open frontier is how you seed an ecosystem, win the developer base, and move the world's default tech stack onto a Chinese one, motivated by technological…
100,000 leads: $7 against $3,000.
Opus 5.5 is 60% cheaper than Astra, at $4 versus $10 per million input tokens.
So far, neither compute nor model advancements have been a zero-sum game: better, cheaper intelligence is accruing value to the entire ecosystem, with both older chips and models retaining substantial value well past the expiration dates the bears…
11 voices
The filing sets the price
Anthropic's expected ~$2 trillion IPO valuation is not justified by its numbers, given that it would need $1.2 trillion in annual revenue within 10 years (per Aswath Damodaran) to justify that valuation while the entire current AI products and…
Anthropic's customer concentration is a risk for its $2 trillion valuation: nearly a quarter of its 2025 revenue came from two customers and 47% of sales ran through Amazon and Google, which is an acceptable problem for a small company but not for…
Distillate Capital Partners LLC
Circular financing distorts the picture: hyperscalers and NVIDIA have committed roughly $180 billion to OpenAI and Anthropic, which pledge about $1.1 trillion of compute spending back to them against far smaller revenues, and $53 billion of…
Alphabet's reported 112% Google Cloud backlog growth is suspect because a new category of unknown size — TPU system sales contracts — was added to cloud backlog in the last six months, and one of the largest TPU customers is Anthropic, in which…
Aggregate purchase commitments, future leases, guarantees backing third-party debt, and various off-balance sheet potential liabilities across the hyperscalers total about $3 trillion, pointing toward a couple more years of record AI data center…
Warden Capital argues the AI investment cycle cannot earn a reasonable return absent AGI within a few years: at roughly $700 billion per year of investment, model providers would need about $1 trillion in annual end AI revenue by 2029, ramping…
Global AI spending is projected to reach 2.59 trillion dollars in 2026, and to achieve even a relatively low unlevered return on investment would require 207 billion dollars in after-tax income per year starting immediately, which the portfolio…
Although SpaceX reports AI infrastructure payback periods of less than one year, maintaining a one-year payback across the entire hypothetical 2027 buildout would require $1.9-2.28 trillion in annual compute payments, and if AI labs footed that…
OpenAI's combined annualized revenue with Anthropic reached $105 billion (OpenAI $40 billion, Anthropic $65 billion) in July 2026, which is nowhere near the figure needed to support the modeled 2027 buildout, and this understates the true gap since…
Anthropic looks likely to file for an IPO soon, and it will want to do that while token usage numbers are still near their peak, before the picture gets less flattering.
The AI trade's foundations are shaky because roughly seventy percent of the hyperscalers' AI revenue traces back to just two customers, OpenAI and Anthropic, both of which are burning enormous amounts of cash, so if either stumbles the entire…
7 voices
Voices that cross sides
OpenAI is in deep trouble in 2027 due to pressure in consumer AI from Meta and SpaceXAI, pressure from Chinese open-source models on cost and token marketshare, and being late to IPO while being overtaken in ARR and growth by Anthropic in…
Michael Spencer argues that in 2026 Chinese open-weight models are taking serious market share from incumbents like OpenAI and Anthropic, because the Chinese labs are only a few months behind the frontier and shortening release gaps mean even…
Cheaper intelligence lifts all; at the model layer no edge sticks.
Oracle's remaining performance obligations of $664 billion are concerning because roughly 50% come from OpenAI, a company burning massive cash, making Oracle over-reliant on OpenAI's solvency.
$70B of ARR at OpenAI; Huang's 80% of AI-native startups on open models.
Chinese models at 66% of tokens; Karp's never-IPO call on OpenAI.
The dollars with two labs, the volume leaving, fragmentation in 3-5 years.
Where they cross
Who answers whom.
The specific points where one voice meets another: the same evidence read two ways, or the same mechanism with a different sign.
Michael Spencer contradicts Tae Kim, on the evidence
Michael Spencer contradicts Le Shrub, on the evidence
Michael Spencer qualifies Wyndo, on a condition
Michael Spencer qualifies Oguz Erkan, on the evidence
Michael Spencer contradicts Le Shrub, on premise
Michael Spencer qualifies Tae Kim, on the evidence
David George qualifies Michael Spencer, on the mechanism
Michael Spencer qualifies Tae Kim, on a condition
David George qualifies Steve Eisman, on the mechanism
Michael Spencer contradicts Wyndo, on the evidence
Michael Spencer contradicts David George, on the mechanism
Kevin Gee contradicts David George, on the mechanism
Tae Kim contradicts Michael Spencer, on premise
Michael Spencer qualifies David George, on the mechanism
Gaetano qualifies Oguz Erkan, on a condition
David George qualifies Michael Spencer, on a condition
Daniel Romero contradicts Market Sentiment, on the evidence
Longriver contradicts Warden Capital, on the evidence
Longriver contradicts First Pacific Advisors, on the evidence
Longriver contradicts Distillate Capital Partners LLC, on the evidence
Michael Burry qualifies Daniel Romero, on the evidence
Michael Burry contradicts Market Sentiment, on the evidence
Gaetano qualifies Mia Silverio, on a condition
Alternative explanations
Four other readings.
What would have to be true for the Call to be wrong, as the other side argues it.
- The labs' price shows in the capex they pull: $1.2 trillion or more of AI capex next year should drive about $2 trillion of incremental revenue, with no prospectus needed (the case Market Sentiment makes).
- The frontier is still paid, and paid more each year: Anthropic's run-rate went from $1 billion at the end of 2024 to $9 billion a year later and almost $50 billion by May 2026 (the case Longriver makes).
- No one commands a price, the lab included: OpenAI's price cut with a tenfold jump in usage looks like the start of a price war between Anthropic, OpenAI and the Chinese models (the case Steve Eisman makes).
- One lab commands the price and the other does not: Anthropic is far ahead of OpenAI in model quality and ARR before its IPO, and the lead should widen with the proceeds (the case Michael Spencer makes).
What settles it
The dated tests.
The tests the story set, each with its specification where the owner has written one and the piece of the argument it verifies; on the Docket, resolved on the Results page.
Alphabet and Microsoft report, and the Fed decides the same day: whether the TPU system sales inside Google Cloud's backlog get a size, and whether Cloud growth holds near the second quarter's 82% without them.
Part 12: The Model Is Not the Price Pendingp12-t01Pre-specifiedChecks the TPU-backlog claim
- Question
- Did Alphabet, in its third-quarter 2026 results of October 28, 2026 (the earnings release, the 10-Q or the call), state a size for the TPU system sales contracts inside Google Cloud's backlog?
- Source
- Alphabet Q3 2026 earnings release, call transcript and Form 10-Q (SEC EDGAR, CIK 1652044)
- Rule
- The answer is yes if the release, the 10-Q or the call transcript gives a dollar amount, or a share of the Cloud backlog, for TPU system sales contracts or for the Anthropic contracts; no if the quarter is reported with no such figure. Google Cloud's growth rate is read alongside (the Part watches whether it holds near the second quarter's 82% without those contracts) but does not decide the test.
- If the source is late, revised or silent
- A size given only as a range wider than two to one, or given for a period other than the backlog at quarter end, is inconclusive, and the Result says what was read and what was missing. A report delayed past November 15, 2026 is inconclusive.
The Call, before the dateExpects no. Gives the yes answer 20%.
0 · no100 · yes
Amazon reports: it does not break Anthropic out, so the test is what AWS discloses about its largest customers, against Silverio's October 5 count of 47% of Anthropic's sales through Amazon and Google.
Part 12: The Model Is Not the Price Pendingp12-t02Pre-specifiedChecks the customer-concentration claim
- Question
- Did Amazon, in its third-quarter 2026 results of October 29, 2026 (the earnings release, the 10-Q or the call), disclose a customer or a group of customers at 10% or more of AWS revenue, or quantify Anthropic's share of AWS sales?
- Source
- Amazon Q3 2026 earnings release, call transcript and Form 10-Q (SEC EDGAR, CIK 1018724)
- Rule
- The answer is yes if the concentration note of the 10-Q, the release or the call names a customer or a group of customers at 10% or more of AWS revenue, or states Anthropic's share of AWS sales; no if the quarter is reported with no such disclosure. The figure is read against Silverio's October 5 count of 47% of Anthropic's sales running through Amazon and Google.
- If the source is late, revised or silent
- A concentration disclosure given only for Amazon as a whole, without an AWS figure, is inconclusive, and the Result says what was read and what was missing. A report delayed past November 15, 2026 is inconclusive.
The Call, before the dateExpects no. Gives the yes answer 15%.
0 · no100 · yes
Anthropic's listing, Part 10's test as the Docket states it: a registration statement with audited revenue, net loss and customer concentration public on EDGAR, or shares trading; a confidential draft without an EDGAR filing is inconclusive.
Part 12: The Model Is Not the Price Pendingp12-t03Pre-specifiedChecks the audited-revenue claim
- Question
- Did Anthropic file a registration statement (Form S-1 or equivalent) with audited financial statements on SEC EDGAR, or begin trading, by October 31, 2026?
- Source
- SEC EDGAR, Anthropic filings
- Rule
- The answer is yes if a registration statement for Anthropic carrying audited revenue, net loss and customer-concentration disclosures is public on EDGAR by October 31, 2026, or its shares begin trading by that date; no otherwise. The audited figures against the $400 billion run-rate line Romero named for the end of 2027, and the two-customer share Silverio cited, are DeepStack's reading into the Result.
- If the source is late, revised or silent
- A confidential draft filing reported in the press but not public on EDGAR, or a filing without audited statements, is inconclusive, and the Result says what was read and what was missing.
Nvidia reports: whether it names the labs among its largest customers, against Eisman's September 21 count of roughly 70% of the hyperscalers' AI revenue from the two.
Part 12: The Model Is Not the Price Pendingp12-t04Pre-specifiedChecks the lab-as-customer claim
- Question
- Did Nvidia, in its third-quarter fiscal 2027 results of November 18, 2026 (the earnings release, the 10-Q or the call), name OpenAI or Anthropic as a customer at 10% or more of revenue, or state the two labs' share of its revenue?
- Source
- Nvidia Q3 FY2027 earnings release, call transcript and Form 10-Q (SEC EDGAR, CIK 1045810)
- Rule
- The answer is yes if the concentration note of the 10-Q, the release or the call names either lab as a direct or indirect customer at 10% or more of revenue, or states a share of revenue for the two labs together; no if the customers at 10% or more are described without naming either lab and no share for the labs is given. The figure is read against Eisman's September 21 count of roughly 70% of the hyperscalers' AI revenue coming from the two labs.
- If the source is late, revised or silent
- A disclosure that names a lab only as an indirect end customer without a share is inconclusive, and the Result says what was read and what was missing. A report delayed past December 15, 2026 is inconclusive.
The Call, before the dateExpects no. Gives the yes answer 15%.
0 · no100 · yes
- When it files
The prospectus: whether the gross margin clears half of revenue, and revenue tied to investors against Distillate's July 15 count of $180 billion committed and $1.1 trillion pledged back.
Part 12: The Model Is Not the Price Pendingp12-t05Pre-specifiedChecks the inference-margin claim
- Question
- Does the first registration statement Anthropic makes public on SEC EDGAR (Form S-1) show a gross margin of 50% or more for the latest period in its statement of operations?
- Source
- SEC EDGAR, Anthropic filings (Form S-1 and its amendments)
- Rule
- Gross margin is revenue less cost of revenue, divided by revenue, from the lines of the statement of operations in the first Form S-1 Anthropic makes public, for the latest period it presents (a 2026 year-to-date period when the filing carries one, else the latest fiscal year). The answer is yes if it is 50.0% or more; no if it is below 50.0%. A margin the filing gives for inference alone, or before payments to cloud partners, is read alongside and does not decide the test; later amendments do not reopen it.
- If the source is late, revised or silent
- If no registration statement for Anthropic is public on EDGAR by March 31, 2027, the test is inconclusive: the Result names the EDGAR search read that day and the missing filing, and the question moves to the first registration statement or annual report Anthropic files, read by June 30, 2027. If the statement of operations carries no cost of revenue line, the gross margin the filing states for the same period decides; if it states none, the test is inconclusive, and the Result says what was read and what was missing.
The Call, before the dateExpects yes. Gives the yes answer 60%.
0 · no100 · yes
- When it files
The prospectus's customer note: whether the customers at a tenth of sales or more add up to a fifth of revenue, against Silverio's October 5 count of a quarter from two customers.
Part 12: The Model Is Not the Price Pendingp12-t06Pre-specified
- Question
- Does the first registration statement Anthropic makes public on SEC EDGAR (Form S-1) report customers at 10% or more of revenue that together account for 20% or more of revenue in the latest fiscal year it presents?
- Source
- SEC EDGAR, Anthropic filings (Form S-1 and its amendments)
- Rule
- Add the shares of revenue the filing gives, in its notes or its risk factors, for each customer at 10% or more of revenue in the latest fiscal year presented. The answer is yes if the sum is 20.0% or more; no if it is below 20.0%, including when the filing states that no customer reached 10% of revenue. A cloud provider that resells the models counts as a customer when the filing counts it as one; later amendments do not reopen the test.
- If the source is late, revised or silent
- If no registration statement for Anthropic is public on EDGAR by March 31, 2027, the test is inconclusive: the Result names the EDGAR search read that day and the missing filing, and the question moves to the first registration statement or annual report Anthropic files, read by June 30, 2027. A filing that names customers at 10% or more without their shares is inconclusive, and the Result says what was read and what was missing.
The Call, before the dateExpects yes. Gives the yes answer 70%.
0 · no100 · yes
What changed
The record of Part 12.
The Part’s publication, every change to the Call, every result that came in and the next dated test, oldest first.
- PublishedPart 12 published; the Call: The price has left the model.
- Call updatedThe Call: The price has left the model. A frontier model no longer commands a price per token for long. The lab still does, and only a prospectus says what that price is worth. Our Call takes the side “The filing sets the price”. by Marco Cauduro
- Next testNext: Oct 28, 2026, Alphabet and Microsoft report, and the Fed decides the same day: whether the TPU system sales inside Google Cloud's backlog get a size, and whether Cloud growth holds near the second quarter's 82% without them. Pre-specified
Sources (16)
- AI Supremacy - Gemini-4 Argon and the OpenRouter token count
- Longriver Investment Partners - Q2 2026 letter
- Prof G Markets - Anthropic wants $2 trillion anyway
- Distillate Capital Partners - Q2 2026 letter
- Capitalist Letters - Booking Holdings deep dive, the token dollars
- a16z - OpenAI understands something important and rare
- AI Supremacy - The era of personal super-intelligent agents
- Apollo - Chinese models vs. frontier models
- Shrubstack - Uncomfortable truths
- Potential Multibaggers - Vercel's open-model count
- a16z - State of Markets II
- First Pacific Advisors - Q2 2026 letter
- Warden Capital - Q2 2026 letter
- Market Sentiment - The next great AI trade is everything that isn't AI
- HyperTech Invest - The math behind the next AI rally
- AI Supremacy - A guide to Grok Bot 2026, the listing delay
Market data are as of the dates cited. Voices are paraphrased from public writing and attributed by name.
Analysis and opinion, not investment advice. Voices are paraphrased from public writing and attributed by name; DeepStack shows no score, ranking or accuracy for any person. Every question, one page each.