Automate Work, Amplify Life.

EAST/WEST AIPHA

Outside China, half the AI revolution is invisible.

Two viral stories, one real rally, a price war settled in the token market.

Edition 20, October 13, 2026. Premium edition.

By San Eng, WSJ bestselling author of The Ten Commandments of Investing.

Hero: Chinese models' share of OpenRouter token volume, February 2026 against the week of September 14, 2026. Source: CNBC, September 26, 2026, reporting OpenRouter data.

落地为王 (luòdì wéi wáng), "implementation is king." Announcements are cheap. What counts is what lands: priced, shipped, and used. This week the West announced. The East landed it.

Price AI by the token, not by the headline. A headline is a claim. A token price is a transaction. This week proved it three times: 722 manuscripts posted, only 162 with a machine-checked proof at launch; Adobe's suite cloned free, the stock higher; 57 to 67 percent of company token volume on OpenRouter runs on Chinese models, depending on the region, at prices 60 to 90 percent lower. Three questions price every AI story: posted or proved, what the stock did close to close, and what the output costs.

COMMANDMENTS #10Cs

Never pay for an announcement. Commandment Five, Seek Religion: find intrinsic value before you pay. A price is a fact. A headline is a claim wearing a fact's clothes. Buy the fact. Let someone else fund the claim.

Post it: OpenAI published 722 manuscripts from about 4,000 posed problems; only 162 carried a machine-checked proof at launch.

Clone it: one developer rebuilt Adobe's suite with Claude. Adobe ended the period higher.

Price it: Chinese models carried 57 to 67 percent of company token volume on OpenRouter, at prices 60 to 90 percent below US frontier prices.

On October 6, OpenAI posted 722 manuscripts, posed about 4,000 open problems, and kept 372 result families, at an average of about three hours of ChatGPT Pro thinking compute per result, by its own repository notes.

Within a day, three manuscripts were withdrawn: a sign error in a paper on Weil classes broke its argument, and two papers built on it came down. None of this is scandal; it is process: publish first, verify later.

At launch the repository catalogue listed only 162 manuscripts with a Lean-formalized main result; the rest had no machine-checked main result and no peer review, and OpenAI warned that unformalized results could contain errors, so posting is not proving. Some will hold. Some will join the withdrawn three.

The capability underneath is real. FrontierMath Tier 4, the hardest public math benchmark, is saturated: top score about 5 percent at its July 2025 launch, 97.6 percent now from GPT-6 Astra on the revised v2 set of 43 problems. Terence Tao had already named the wider shift proof indigestion: AI can generate proofs faster than the community can verify, explain, and canonicalize them. Adjudication is the scarce product.

Both things are true: the math is real, and nothing was "solved" until a machine or referee checked it.

Chart 1: the manuscript funnel. Source: OpenAI repository catalogue (lean/formalization.yaml), October 6, 2026; Retraction Watch, October 8, 2026. Chart 2: FrontierMath Tier 4, 5 percent to 97.6 percent. Source: Epoch AI, October 2026.

The builder is one developer, Brandon Thomas of Atlanta, founder of ArtCraft and a former Square engineer, who led the build and took community contributors on GitHub after launch; the four-person Atlanta team line in IT Home and 17173 traces only to Futurism and is corroborated nowhere. Working with Claude Opus 5.5, credited as co-author on most of PhotoCraft's commits, 757 of 1,207, about 63 percent, in a GitHub API count dated October 10, 2026. He published seven Adobe-facing apps: PhotoCraft, VectorCraft, FilmCraft, LightCraft, EffectCraft, DesignCraft, and PrintCraft, since renamed PdfCraft, which the project describes, without independent audit, as clean-room rebuilds from public specifications and observed behavior. Word, Excel, and PowerPoint counterparts followed on October 7.

The clones are real and rough. PhotoCraft passed 50,000 GitHub stars and about 8,100 forks by October 12, up from 40,459 on October 10, and above 53,500 a day later. Thomas walked back full parity: 99 percent will take months. In PhotoCraft's own parity estimate of October 3, 2026, about 111 of its 170-file PSD test corpus, about 65 percent, pixel-matched Photoshop in the project's own oracle test, a self-assessed figure its roadmap raised to 115 of 170, about 68 percent, on October 5. One roadmap concedes the cloud layer: PrintCraft's roadmap excludes Adobe's Document Cloud storage, Adobe Sign, and the Acrobat AI Assistant from its parity target, which have no clean-room equivalent. That assistant sits inside Adobe's fastest-growing product line, AI-first ARR above $500M and tripled year on year.

Adobe rose 3.56 percent on October 8, closing at $241.05 against $232.77 on October 7. The stock ended higher, not lower; the crash was a headline in search of a chart.

Chart 3: Adobe daily closes across the clone window. Source: Yahoo Finance daily closes, retrieved October 12, 2026.

So why no crash, if writing software costs almost nothing? Because software is four assets wearing one label.

The "software is dead" canon: Andreessen in 2011, Huang in 2017, Nadella in 2024 calling business apps databases. Is this time different? In one way, yes: the copy costs nothing and arrives in weeks, so the fourth class has nowhere to hide. In the way that decides the first three, no: every wave killed the same layer, the one with no workflow around the code. Code was never the moat. The workflow is: files, plug-ins, teams, trust.

While the West argued headlines, developers rerouted traffic. CNBC reported Chinese models carried 57 to 67 percent of company token volume on OpenRouter in the week of September 14, depending on region, up from 6 to 13 percent in February. The reason, per OpenRouter data shared with CNBC: Chinese open models do agentic work, at prices 60 to 90 percent below US frontier models. Chinese coverage reads the price war as deliberate strategy: commoditize the model layer, push value up into applications and distribution.

Chart 4: Chinese share of OpenRouter token volume. Source: CNBC, September 26, 2026, reporting OpenRouter and Vercel data.

Tokens are traffic. Revenue is money. Token share measures who is being tried; revenue share measures who is being paid. Hold the distinction and the scope: by a tally relayed by BlockBeats and republished by KuCoin on October 8, 2026, attributed to Social Capital using OpenRouter data, DeepSeek (深度求索) models processed 36.8 trillion tokens on OpenRouter in the last week of September, against 33.0 trillion on that platform for OpenAI, Google, Anthropic, and xAI models combined. Lindy, the San Francisco AI-employee startup, is the named case: CNBC reported on June 26, 2026 that its monthly AI bill had passed its total salary spend and that it had switched all traffic to DeepSeek, and Atlas Cloud later reported that Lindy moved most of its managed-agent traffic off Claude and Gemini onto DeepSeek v4 Flash on Atlas Cloud, cutting inference costs on the migrated routes by about 90 percent.

Chart 5: the Lindy switch, inference cost on migrated routes, indexed. Company-reported figures. Source: Atlas Cloud case study, 2026; CNBC, June 26, 2026.

The capital is following. Bloomberg reported on October 6, 2026 that DeepSeek was close to securing at least 80B yuan, about $11B to $12B, in a round that could reach 100B yuan, with Tencent and CATL among the largest committers; the 500B yuan figure was the valuation DeepSeek targeted when it launched the raise in July. Moonshot (月之暗面), maker of Kimi, closed its final private round near a $50B valuation; annual recurring revenue ran from $300M in June to about $1B now, and a Hong Kong IPO is targeted for the first quarter of 2027.

Chart 6: Moonshot valuation against annual recurring revenue. Source: Bloomberg, October 6, 2026.

China is not winning every layer. FrontierMath Tier 4 (the v2 set) problems are so hard that solving one is a publishable result. Astra's 97.6 percent means the frontier model closes them at will; the roughly 2 percent from the best Chinese models leaves research-grade work out of reach. Commodity work goes to the cheapest model that can do it; the hardest work belongs to the frontier, and the frontier is American. On open mathematics, China moved first: DeepSeek-Math-V2, open-sourced in November 2025 under Apache 2.0, scored at gold-medal level on IMO 2025 problems, five of six solved.

The West leads in capability. China leads in commoditization. Leading in capability means doing the work nobody else's can yet; in commoditization, doing the work everybody needs, at a price nobody can match, on rails everybody uses. On formal proving, China co-leads: DeepSeek-Prover-V2 88.9 percent, Kimina-Prover 92.2 percent, ByteDance Seed-Prover 99.6 percent on miniF2F.

The math breakthrough and the clone wave are American. China's press read the clones plainly: rebuilding software from outside is an old technique; what is new is one developer directing AI agents. China's contested shortcut is different: distillation, training smaller models on frontier output, which Anthropic alleges seven labs ran at scale.

Where China is ahead is the rail, built top-down. In February 2025, in the weeks after R1 launched on January 20, Huawei, Honor, OPPO, and Xiaomi announced or rolled out DeepSeek integrations, vivo announced a planned integration, and Meizu said its Flyme AIOS completed integration that month. The three state carriers integrated it across their networks in front of about 1.7 billion mobile subscribers. Adoption decided at the top buys instant reach, and reach generates the data. Apple Intelligence reached China after regulators registered it this July, running on Alibaba's (阿里巴巴) Qwen (千问), an integration both have confirmed, rolling out. Distribution in China is the default setting on the device in your pocket.

The strongest case against today's framing. One: posted is not proved cuts both ways; the 97.6 percent Tier 4 score is a frontier signal, and three withdrawals are a rounding error. Two: clone quality compounds; version two may not stall, and moats break. Three: the probe may be about controlling the IPO calendar, not data; state support usually arrives before state punishment.

For a limited time, the full premium edition is free for subscribers.

Seven of our ten seats land in the same place: if software costs nothing to copy, durable exposure is the permission and billing layer, Big Tech distribution and the cloud. Nine of ten would not sell Adobe-class incumbents into this fear. For the full list of the 10 value investors, see the Reference section below.

Analysis, not advice.

BUY: Big Tech distribution owners: Apple, Alphabet, Microsoft, Meta, Tencent. These businesses own the moat: trust, workflow, distribution, permission. Lowest-cost AI infrastructure and hyperscale cloud: when a commodity's price collapses, own the mine and the railroad. China open-weight distribution at a discount: operators with enterprise users and real throughput, never the weights.

HOLD: Software incumbents with deep workflow lock-in, the Adobe class. The clone week did not sink Adobe; it closed higher; selling proven cash flow into narrative fear is the temperament error. Frontier labs in either country: capability is real; which management survives commoditization, no seat can yet say.

SELL: Seat-license software with thin workflow lock-in; zero marginal cost reprices it first. Prompt wrappers and no-moat agent startups at hype valuations: zero cost plus zero moat equals zero pricing power. Pure China open-weight exposure with no paid enterprise distribution: free scale without a path to paid is research, not a business.

Post less. Prove more. Clone less. Distribute more. Price everything by the token.

That is Commandment Five in action: pay for what is verified, own what cannot be copied cheaply, and let the other man's headline carry his risk.

Reply and tell us: the headline, or the token?

San Eng

Chairman, WorkOptional.ai

This newsletter is investing education, not investment advice. Nothing here is a recommendation to buy or sell a security. When we hold a position in a company we discuss, we will say so.

This newsletter is San Eng's, written by San, WorkOptional's Chief Investment Officer, and researched with his team. It doubles as the research behind his next book, The Ten Commandments of Investing in the Age of AI, published as it happens. The team: Alvin Foo, Chief Marketing Officer of WorkOptional and former Head of Mobile, Greater China, at Google, makes the research travel. Dr Felix Arndt, Professor at the University of Waterloo and Director of the Conrad School of Entrepreneurship and Business, makes the ten commandments hold up to scrutiny. The full archive lives on the WorkOptional site.

This is a virtual investment board that learns from this group of 10 long-term value investors. We asked each seat the same question: if software costs nothing to copy, what is left to own?

The ten lenses: 1. Buffett: owner earnings; the moat is distribution plus trust. 2. Grantham: valuation discipline; margins revert. 3. Lynch: buy what you understand, at a reasonable price. 4. Baron: hold for decades; compounding is the edge. 5. Pabrai: clone the best ideas openly; few bets, held long. 6. Zhang Lei: structural demand; bet where the trend is irreversible. 7. Li Lu: price versus value; margin of safety is everything. 8. Duan Yongping: pricing power in a bad year. 9. Kathy Xu: back the founder and the distribution. 10. Ho Ching: duration matching; patient capital for long assets.

What would change our minds: agent unit economics flipping below seat cost at scale, a Big Tech drawdown of 20 percent or more on fear rather than fundamentals, Chinese enterprise monetization overtaking free consumer volume, or Meta opening, taxing, or closing agent distribution on WhatsApp.