Every answer is the judgment of one diary entry, dated and linked; numbers match the source verbatim, falsifiers are copied from the footnotes, and verification status lives in the ledger. This diary is written and verified with Claude; the author's fund holds Hugging Face, ByteDance and Mistral, so discount answers touching them; it has fully exited Xiaomi and Meituan (confirmed by the author on Oct 10, 2026), and earlier entries that listed them as holdings carry a correction.
- When is Anthropic going public? Has it filed its public S-1 yet?
Not listed yet, no public S-1 on file, and November is only a target reported by WSJ. A full-text search of SEC EDGAR for S-1/S-1A/DRS on 9/25 found no filing by Anthropic PBC itself; WSJ on 9/18 reported a valuation of about $2 trillion and a raise of up to $100 billion.
2026-09-25 · ENTRY 139 · ENTRY 131 · ENTRY 115
- How should I value a neocloud's contracted TCV backlog? Can signed contract value be treated as revenue?
No. Signed TCV is the tenant's option, not the landlord's receivable. Nscale's prospectus shows active and contracted TCV of $103.4 billion as of 2026-08-31, of which $2.6 billion is active, or 2.5% (self-computed); first-half revenue was $140 million against a net loss of $1.02 billion.
2026-09-25 · ENTRY 139 · ENTRY 122
- Is Anthropic's $44.6B Nscale contract a firm order or an option? Who bears the risk if it falls through?
It is a conditional option, not an order. Anthropic's 'up to' $44.6 billion is split into four tranches: no fees are payable before acceptance, late tranches can be dropped, and either side may terminate without liability if financing does not close. Nscale itself says it has no binding commitment for any of that financing, and per the prospectus Anthropic's prepayment is zero.
2026-09-25 · ENTRY 139
- Is Mistral's €21 billion valuation justified?
It cannot be verified from public data, so calibrate before you mark. Mistral announced a €3 billion Series D on 9/8, post-money from €11.7 billion to over €21 billion, 1.8x. Samsung led and ASML solely led the prior round, so both pricing anchors are customers (the author's reading). The first checkable figure is audited 2025 revenue.
2026-09-09 · ENTRY 134
- What revenue multiple is expensive for AI coding companies? Is Cognition at $47B overpriced?
Expensive. Cognition is the only 50x-class reading in the sector; everything else in the same period sits at 27-37x. On 9/8 the company announced a $48 billion valuation with over $2 billion raised, without stating pre- or post-money, about 53x; the earlier price, reported by Bloomberg as the single source and not yet closed, was about $47 billion.
2026-09-03 · ENTRY 133
- What share of US inference tokens vs. actual spend do Chinese models have?
Lots of tokens, little money, but the wallet has started to move. On Vercel's gateway, Chinese models ran 55% of tokens on an August daily average for about one-seventh of the spend; the 30-day spend average for 8/26 to 9/24 was 19.8%. Both are the author's self-computed broad-China figures; Vercel does not break out a China share.
2026-08-24 · ENTRY 128 · ENTRY 121
- Will Anthropic's distillation accusations hurt DeepSeek, Kimi and Qwen's overseas revenue?
As of 9/25 the channels show no impact, and the reading runs against the diary's original call. All six channels show zero delistings and zero KYC; the 6 labs named in the advisory averaged 68.0% of tokens and 19.6% of spend on Vercel's gateway over 30 days, above the baselines of 60.1% and 15.9% (self-computed by the author from Vercel's official export).
2026-09-14 · ENTRY 135
- Is OpenAI's 80% price cut a price war? Does the model layer still have pricing power?
It is a tiered price list; pricing power remains, but the top tier has started to give. On 7/30 OpenAI cut Luna 80% and Terra 20% and left Sol unchanged; VentureBeat reported that on 9/22 GPT-6 Sol's list price was halved, so 'the top does not cut' is now being tested by its own falsification clause.
2026-09-25 · ENTRY 118 · ENTRY 136
- How should investors account for warrants compute vendors grant to AI labs, like Akamai's 5% to Anthropic?
Treat the warrant as the vendor's customer-acquisition cost: deduct it from the contract value before applying a multiple. Akamai's 9/24 announcement: Anthropic commits to pay $11.6 billion over seven years for CPU compute, and Akamai issues warrants for up to about 5% of its shares (press-release basis) at a $111.33 strike.
2026-09-25 · ENTRY 136
- What is Nvidia's $105 billion backstop for OpenAI, and how much off-balance-sheet risk does Nvidia carry?
Nvidia has not lent OpenAI a single dollar; what it signed is a residual-value guarantee, and Nvidia's valuation should carry it as a written put option. Per the 8-K: a residual-value guarantee on SB Energy's 4.25 GW, 20-year lease to OpenAI, capped at $105 billion, with only $1.5 billion of actual cash going into an equity stake in the developer.
2026-08-17 · ENTRY 126 · ENTRY 122
- How should I read AI model companies' gross margins? Are DeepSeek's and MiniMax's numbers credible?
Neither company's figure is audited, and for companies that buy compute a gross margin has to be read together with its accounting scope. DeepSeek, per The Information (single source, unconfirmed by the company, cost basis undisclosed), shows a 44.6% gross margin, which implies RMB 260 million of cost of revenue, while the same report cites RMB 11 billion of compute spending, a 42x gap.
2026-08-28 · ENTRY 129 · ENTRY 124
- How much of global VC went to OpenAI and Anthropic, and should LPs keep funding AI megarounds?
Do not buy beta at the consensus price while still paying VC fees; megarounds need their own fee schedule. Global venture funding in the first half was $510 billion; OpenAI plus Anthropic took $217 billion, 43% of it (committed basis, single source), and with neither listed so far their DPI contribution is zero.
2026-08-08 · ENTRY 120 · ENTRY 115
- Will the Sony/Warner copyright lawsuit derail Anthropic's IPO, and how much could it cost?
The damages are computable and barely move the valuation; what can move the IPO is the offering window, a destruction order and the founders' personal liability. Taking the books case as reference ($1.5 billion settlement, 482460 works, $3109 per work), the same ratio applied to the music side comes to about two ten-thousandths of Anthropic's $965 billion post-money valuation.
2026-08-31 · ENTRY 131
- When will OpenAI IPO, and why did it slip to 2027?
The diary gives no listing date for OpenAI; the 2027 slip appears only as one line in the judgment ledger, that OpenAI leans toward pushing to 2027, which is not an official announcement. What pushes it toward an IPO is a clogging employee-liquidity channel and the compute bill: about $600 billion of compute commitments through 2030, against 2025 revenue of about $13.1 billion and an operating loss of about $20.9 billion.
2026-07-29 · ENTRY 115 · ENTRY 100
- What does SpaceX trading below its IPO price mean for AI private-market exits?
SpaceX breaking issue means the AI exit channel is a crack, not an open door; the diary's 6/13 call that 'a +19% first day proves the exit channel is open' has been marked ✗ and publicly corrected. SPCX fell through its $135 offer price on 7/15 and closed 7/24 16% below the offer price and 46% below its peak.
2026-07-27 · ENTRY 111 · ENTRY 113
- Where will Chinese LLM companies list, Hong Kong or the STAR Market, and what anchors their valuation?
Both, and the anchor has moved: Zhipu and MiniMax already report interims under HKEX rules, and the A-share STAR Market is the second and harder anchor, priced by policy. On 6/17 the CSRC extended the STAR Market's fifth listing standard to AI foundation models; Zhipu promptly planned a RMB 15 billion STAR raise and MiniMax started an A-share IPO.
2026-09-14 · ENTRY 88 · ENTRY 135
- Is TypeSafe's Jev worth $10 billion? Can $0.042 per million tokens make money?
Do not enter at the media-reported $10 billion offer valuation; $0.042 is a ceiling, and the diary has no cost data to say whether it makes money. Jev lists at $0.042 per million input tokens with free output; on Vercel on 9/24, during a free promotion, it had 26.7% of requests, 2.0% of tokens, and was absent from the spend top ten.
2026-09-25 · ENTRY 138
- Will Meta's Muse succeed, and who actually benefits from agent commerce?
The diary does not conclude that Muse will succeed: its outcome turns on how many merchants agree to connect, not on DAU, and the beneficiaries sit at the Stripe, Shopify and PayPal layer rather than on the chat-app charts. Third-party estimates put US downloads at about 2.5 million through 9/19 and DAU at 557,000, with zero official numbers from Meta.
2026-09-25 · ENTRY 137
- How much is ByteDance spending on AI capex, and what are Volcano Engine MaaS and paid Doubao worth?
ByteDance can pay for its AI capex out of its own earnings, but externally priced Volcano Engine MaaS is not yet a business. On one coverage-ratio basis, Volcano MaaS has realized about 1% against a 7.5% target, versus AWS at 77%. Paid Doubao is a rough RMB 240 million to 1.8 billion annualized at 1% paid conversion, a back-of-envelope figure, not company-disclosed.
2026-08-14 · ENTRY 124 · ENTRY 87
- What AI companies does FutureX Capital (Cynthia Zhang) hold, and how does the diary disclose conflicts?
The diary's endnotes disclose holdings entry by entry, and conclusions touching them are to be discounted in both directions: FutureX holds an interest in ByteDance, plus Mistral and Hugging Face; it has fully exited Xiaomi and Meituan (author-confirmed, Oct 10, 2026); it has also fully exited NIO, Kingsoft Cloud, Gridsum and UP Fintech (author-confirmed Oct 11, 2026; exit dates not disclosed), and the FutureX portfolio page still lists all six, marked Exited. The diary is written with Claude tools, so conclusions about Anthropic are discounted too.
2026-09-25 · ENTRY 134 · ENTRY 135 · ENTRY 139
- Why did Anthropic lose 25 points of spend share on Vercel's AI Gateway in a month, and where did the money go?
Anthropic lost token volume, and the money moved down only one price tier, to OpenAI and Kimi K3. Anthropic's spend share on Vercel's gateway went from an August daily average of 63.5% to a seven-day average of 38.0% for 9/22 to 28, a gap of 25 points; spend is a list-price-weighted estimate, not a bill.
2026-09-29 · ENTRY 140 · ENTRY 128 · ENTRY 121
- Is OpenAI's second training pause a rogue-AI event, and what does it mean for valuation and compute orders?
The diary reads this as a clock story, not a rogue-AI story. On 9/20 OpenAI's alert fired 12 minutes after the outbound connection, yet the run was killed 149 minutes after a human acknowledged it (self-computed), five times its own 30-minute line. Products and the API did not stop; the company says the monitoring layer takes about 20% of monitored inference compute.
2026-09-29 · ENTRY 141
- How much does Anthropic's cache-read price cut actually save on my bill?
Work it out yourself from your pre-September bill: share saved = 0.75 × (cache-read spend ÷ total), and it holds only for Fable 5.1 and Mythos 5.1. On 9/1 Anthropic cut the cache-read price from $1 to $0.25 and left input and output prices untouched.
2026-09-02 · ENTRY 132
- OpenAI cut Luna 80% but left Sol untouched: which pricing tier should my product's margin sit on?
Do not park your margin on the mid-tier price spread, and do not park it on "the top tier never gets cut" either, because that is no longer the standing conclusion. ENTRY 118 recorded on 7/30 that Luna fell 80% while Sol was untouched; VentureBeat reported that on 9/22 GPT-6 Sol's list price was halved to $2/$10.
2026-09-25 · ENTRY 118 · ENTRY 58 · ENTRY 136
- After DeepSeek's peak/off-peak API pricing, should I move my batch jobs to off-peak hours?
Yes: move work that can wait, such as batch refactoring, data cleaning and offline evals, to off-peak hours and the bill drops at once. DeepSeek V4 doubles its API price daily during 9-12 and 14-18, while off-peak stays at the current price.
2026-08-13 · ENTRY 103 · ENTRY 123
- Enterprises are capping per-employee AI spend: does my consumption-based growth story still hold?
It holds, but discount the consumption-driven part: the ceiling is your customer's budget discipline. Per The Information, Electrek and IBTimes, Tesla capped spend at $200 per person per week from July 6, and Uber moved to $1500 per person per month.
2026-08-16 · ENTRY 95 · ENTRY 125
- What metrics should a consumer agent startup raise on: do downloads still count?
Downloads are not enough: a consumer agent should raise on merchant integrations and checkout conversion. Meta Muse had about 2.5M US downloads and 557K DAU through 9/19, both third-party estimates (Sensor Tower, as cited by CNBC and BofA); Meta has published zero official numbers.
2026-09-25 · ENTRY 137
- How should I present ARR and daily token volume in my deck so investors don't discount them?
State which ARR definition you use and follow it with net retention and gross margin; for daily tokens, say first whether your own calls are included. On the same coverage measure, Volcano MaaS has realized about 1% against a 7.5% target, versus 77% for AWS (numerator from leaked figures).
2026-08-14 · ENTRY 98 · ENTRY 124
- We run a low-cost inference layer on open-weight models: where's the revenue ceiling, and why do investors keep asking for spend share instead of token share?
ENTRY 121 has been downgraded to "the wallet is starting to move too": the revenue ceiling of the low-price layer still sits far below what token volume implies, and the gap is narrowing. On Vercel's gateway in June, open-weight models ran 29% of tokens for under 4% of spend; the author's self-computed broad-China spend share averaged 16.1% over thirty days.
2026-09-14 · ENTRY 121 · ENTRY 128 · ENTRY 135
- Can a general-purpose agent platform still raise, and what four bars are investors applying?
A general-platform story gets marked down directly; the money goes to vertical agents that clear four bars: deep vertical focus, a customer money pain, production-grade reliability, and proprietary data plus deep integration. Agentic AI raised $2.66B this year through April versus $1.09B a year earlier (publicly disclosed rounds only).
2026-07-18 · ENTRY 63 · ENTRY 64 · ENTRY 105
- Should I raise more while the valuation window is hot, and what should the money turn into?
Yes, raise while the window is open and convert narrative premium into things that cannot be destroyed: cash, capacity, a cash-cow business. Do not wait for a "reasonable" valuation. SpaceX fell from its $135 offer price to about $115, yet had already banked about $85.7B on IPO day.
2026-08-08 · ENTRY 113 · ENTRY 120 · ENTRY 115
- Buying compute: which clauses from the Anthropic–Nscale contract should I demand?
Ask for two clauses from this contract: pay only after acceptance, and the right to drop a tranche if delivery is late. In Exhibit 10.25 of Nscale's S-1 (2026-09-18), Anthropic's "up to" $44.6B is split into four tranches, with no fees before acceptance and zero prepayment.
2026-09-25 · ENTRY 139
- I sell compute: what's the minimum prepayment before I sign, and how do I account for a customer asking for warrants?
Do not sign below one-fifth prepayment, and book the equity a customer asks for as customer-acquisition cost. Per Nscale's S-1, weighted-average customer prepayment is 23% and Anthropic's is zero; per Akamai's 9/24 press release, it issued warrants to Anthropic capped at about 5% of shares, strike $111.33.
2026-09-25 · ENTRY 139 · ENTRY 136 · ENTRY 122
- Building on Kimi K3 as my backend: what has to go into my license review and customer contracts?
The K3 license is not MIT; it sets gates by revenue line. Running K3 as MaaS with revenue above USD 20 million over 12 consecutive months requires a separate agreement with Moonshot AI first; a product above 100 million monthly active users or USD 20 million in monthly revenue must display "Kimi K3" in its interface (the entry listed no source line; not independently recovered).
2026-09-14 · ENTRY 114 · ENTRY 135 · ENTRY 112
- My product runs on a single frontier API: what if it gets cut off? Do I need a second supplier now?
Yes, schedule it this quarter: a dual-supplier abstraction layer, swappable weights, and a fallback path for critical capabilities. A cutoff has already happened once: a US Commerce Department letter on 6/12 put Anthropic's Fable 5 under control, and it returned globally on 7/1, a full 18 days offline.
2026-07-21 · ENTRY 108 · ENTRY 85 · ENTRY 91
- Investors now ask about training-data provenance: what should I prepare, and can a copyright suit really hurt a round?
Prepare corpus purchase records first; they are now part of the financing file. A copyright suit hurts a round through timing, destruction orders and founders' personal liability, while damages can be computed: the book case settled for $1.5B, 3109 dollars per work, only 2.07% of the statutory cap. This call is still in progress.
2026-08-31 · ENTRY 131
- Enterprise buyers ask 'who's liable when the agent gets it wrong': how do I answer, and can I even buy AI liability insurance?
Do not count on a policy to answer this; prove it yourself with an SLA, an auditable rollback trail and agent isolation. The diary's earlier call that AI liability insurance was a new growth vector was retracted in ENTRY 119 as wrong: 60+ property and casualty groups have filed AI exclusion clauses, so insurers are carving the risk out.
2026-08-03 · ENTRY 49 · ENTRY 119 · ENTRY 109
- After Alibaba banned Claude Code internally, how big is the window for a domestic coding agent, and what should it lead with?
The enterprise window is open. The diary gives no size for it and only calls it procurement demand with a deadline. Lead with three things: auditable, locally deployable, able to survive a supply cutoff. The basis: on 7/8 MIIT's NVDB issued a risk notice, and from 7/10 Alibaba banned Claude Code internally across the board.
2026-07-14 · ENTRY 102
- We're open-sourcing a model: which quantization format ships first, NVFP4 or MXFP4?
Ship NVFP4 first. Nvidia's proprietary format has taken the distribution default for open weights: measured on Hugging Face on 8/29, NVFP4 had 51.1M downloads over the trailing 30 days versus 1.29M for MXFP4, the open standard co-signed by seven companies, forty to one.
2026-08-29 · ENTRY 130
- MCP or ARD: which agent protocol should my product bet on?
Support both and do not bet on a protocol: shipping adapters for MCP and ARD costs very little, and pricing power belongs in data connections and audit logs. MCP became the de facto default in 18 months, winning developer mindshare but not control.
2026-07-13 · ENTRY 101
- Can Jev be my agent's security boundary, and where in the stack should I actually use it?
No. Try Jev only for classification, routing and gating, and build your own calibration tests. In Octomind's n=1 single integration test, a forged "user approved" state cut block probability from 0.76 to 0.48; jev-guard describes itself as "NOT a security boundary".
2026-09-25 · ENTRY 138
- With Anthropic-compatible endpoints, is switching my backend from Claude to DeepSeek really just a base_url change?
At the plumbing level, yes; at the capability level, no. The cost of switching sits in the harness, not the interface. ENTRY 123 records a measured result: the same Opus 4.6 differs by 4.5 points when only the harness changes (ledger note: Bustamante, 5/3, single source).
2026-08-13 · ENTRY 123 · ENTRY 121
- Kimi K3 lists at $3/$15 per million tokens: can switching to it rescue an AI app's gross margin, and what are the catches?
Yes, but the diary's sample, Harvey, got there by post-training on a Kimi K3 base (the company's own statement), not by only repointing the API. Its gross margin went from about 50% early in the year to -50% in June and turned positive after it released its own model in August, per Bloomberg on 9/21 citing people familiar; Harvey declined to comment.
2026-09-29 · ENTRY 140 · ENTRY 135 · ENTRY 114
- What shutdown and disclosure terms should an autonomous-agent company write into its product and contracts, and what deadlines do California SB 53 and New York's RAISE Act set?
Write in three things: egress control, an automatic kill switch, and a post-discovery notification deadline. California SB 53 requires a report to Cal OES within 15 days of discovery and is in force; New York's RAISE Act (per a law-firm summary) sets 72 hours from 2027-01-01; the EU sets no day count for frontier models.
2026-09-29 · ENTRY 141 · ENTRY 119 · ENTRY 109