Market Overview, February 10, 2026

Summary

• Macro verdict: this is not stagflation but a “deflationary boom.” Jin coined the term to push back against the market’s stagnation narrative: “Growth is not bad, corporate earnings are very good, but the jobs are gone.” Weak employment, solid growth, and low inflation are this year’s playbook; last year’s inflation was mainly driven by Trump tariffs, and excluding them, “the risk of deflation in the U.S. is greater than the risk of inflation.” China is already deflationary.
• Be careful with Friday’s payrolls report. ADP, initial claims, and job openings all weakened last week, so payrolls could disappoint. The market has not fully digested the QT expectations triggered by Warsh’s likely appointment and may initially price in “bad data is good data,” but “it could quickly flip to bad data is bad data, and that probability is high.”
• AI is really eating people’s jobs, not SaaS. The U.S. labor market has entered a “low hire, low fire” regime: Goldman Sachs already uses Anthropic for client onboarding and KYC back-office work, where base pay is roughly $200,000 and VP compensation roughly $300,000. The business is expanding, but headcount is not. Of the world’s 700M white-collar workers earning an average $25,000, AI could replace 30%: “New things always have to kill the old before they can create new productive forces.”
• Japan: Takaichi wins by a landslide, and the carry trade continues. She secured 316 lower-house votes, or two-thirds, meaning “she can do whatever she wants and amend the constitution.” Aggressive fiscal policy paired with “irresponsible monetary policy” will push the yen lower against the dollar. Much of last Friday’s and Monday’s U.S. equity rebound was “the carry trade adding back to positions.”
• Two crowded positions are the market’s central risk. One is long the Google basket and short the OpenAI chain—NVDA, Oracle, Microsoft, and Coview (likely CoreWeave). Jin sees no logic in shorting NVDA because everyone needs its compute. The other is short software and long “Sanmi” (storage, with MU the largest ETF weight). If either trade reverses, the unwind could look “very much like the unwind when people bought AMD and shorted Intel.”
• The decisive battleground in the model war is mid-training. Google puts more than 70% of its compute into pre-training general-purpose foundation models; at the time, the gap between TPU and GPU was not that large, and aside from Codex there were no visible models trained on B-series GPUs. DeepMind “has no interest in commercialization,” making it difficult for Google to preserve a general-model advantage over the next year.
• Software is turning from a drug company into a factory, and Nvidia is the “land god.” The old model sold near-zero-cost copies; the new model sells tokens from factories with structurally insufficient supply. Nvidia does not sell cards, only large-system solutions. HBM has been “sold out by Huang”; there is currently a shortage of 3M cards, while datacenters scheduled for completion this year need another 4M. “Every pit Nvidia steps into today becomes a barrier to entry tomorrow.” Jin also likes Grok for its ability to deploy compute clusters and Elon’s product instincts; its only problem is that it cannot get enough cards. Meta and Microsoft will inevitably join the fight because the cloud rentier model has been disrupted.

Deep dive

1. Deflationary boom: deflation in jobs, boom in the economy

• Jin’s core macro call is that the market is fixated on stagflation because “history keeps circling in our heads,” but the actual state should be called a deflationary boom: “Growth is not bad, corporate earnings are very good, but the jobs are gone. When the jobs disappear, costs fall and corporate profits improve further.” Last year’s inflation was mainly caused by tariffs, pressure is limited this year, and ISM shows that business conditions are not poor.
• The data setup is deteriorating: ADP, initial jobless claims, and job openings all weakened last week, so Friday’s payrolls could be very weak. QT expectations tied to Warsh’s appointment have not been fully absorbed. The market may initially price in “bad data is good data,” but “it could quickly flip to bad data is bad data, and that probability is high.”
• The labor market’s structural disease is “low hire, low fire”: companies are expanding without hiring, and while testing AI they are not yet laying off staff. Goldman Sachs is using Anthropic for client onboarding, KYC, and other back-office work—a role paying roughly $200,000 at the base level and roughly $300,000 for a VP—and “when the business grows, headcount will not grow.”

2. Takaichi’s landslide: the carry trade continues under irresponsible monetary policy

• Takaichi won 316 lower-house votes, or two-thirds, while her influence inside the LDP is also overwhelming: “She faces no resistance in executing any policy. She can do whatever she wants. She can amend the constitution.” She won by being bold and therefore will not turn conservative: expect expansionary fiscal policy, a rightward tilt on immigration, and “new capitalism.”
• The market implication is unambiguously risk-on in the short term. “Under irresponsible fiscal policy, monetary policy will not be especially responsible.” Ten-year yields and inflation expectations are rising, and the Nikkei rose 5% yesterday. The yen will depreciate against the dollar, the carry trade will continue, and much of the U.S. equity rebound on Friday and Monday was “the carry trade adding back to positions.”

3. Two crowded positions: long Google/short the OpenAI chain, short software/long “Sanmi”

• Hedge funds have accumulated large short positions since January, but the index has not been pushed lower because the market is trading long/short: long a Google basket and short the OpenAI chain—NVDA, Oracle, Microsoft, and Coview (likely CoreWeave). The logic is that Gemini will rule the market while OpenAI’s weak cash flow will force it into bankruptcy. Jin’s rebuttal is that shorting NVDA “makes absolutely no sense because everyone needs its compute.” Shorting Oracle also makes little sense because it can sell cards without OpenAI.
• The second trade is short software and long “Sanmi”—storage, with MU carrying the largest ETF weight. The catalyst was the release of Claude 4.6 and OpenAI Codex. Jin calls himself “a programming novice” yet wrote an international chess program in minutes, leading the market to conclude that the several-trillion-dollar SaaS market would be replaced. Software rebounded 3% yesterday while “Sanmi” was sold, which is why MU did not rise.
• The risk is a wholesale unwind if either trade reverses: “The effect is actually very similar to the unwind when people bought AMD and then shorted Intel.” The short-software position “will definitely come back, but not necessarily now,” because software cannot currently be valued: 30x EV or a few turns of EV, “there is no answer.”

4. The model war: training is resource allocation, and the decisive edge is mid-training

• Jin’s first-principles framework is simple: “When you train a model, you are fundamentally allocating resources.” Gemini’s performance last year showed that scaling laws still work. Top technical talent returned to repair scaling, while Google leveraged the Android ecosystem—Qualcomm and MTK “are all its little brothers”—to secure TSMC CoWoS capacity and stack large numbers of TPUs. But the TPU-GPU gap was not that large at the time: aside from Codex, there were no visible models trained on B-series GPUs. TPU refresh cycles are roughly 1.5 years versus roughly one year for GPUs. “Whether Google’s models can gain the same advantage as the first two in general-purpose models over the next year is, in my view, very difficult.”
• The strategic divide is clear: Google puts more than 70% of its compute into pre-training general-purpose foundation models, while OpenAI and Anthropic put substantial compute into mid-training for vertical applications. GPT-5.3 was released before mid-training was complete under pressure from Gemini, while GPT-5.2 itself had training problems; after mid-training, 5.3 performed very well in finance and other verticals. OpenAI has enormous cash to buy data: it spent $35M labeling investment-banking IBD data. “If $35M isn’t enough for a lawyer’s case, how about $350M?” Anthropic has fewer resources—roughly 350B in its last financing round—but has cut exceptionally well into programming and is also moving into finance. “That is pretty impressive.”
• Citing Jason Huang (likely Jensen Huang) last Friday, Jin said AI has reached an inflection point: it is moving from something entertaining to something practically usable that can replace labor. The test is “whether someone directly uses every token produced, and whether that token usage can continue.” OpenAI’s January revenue rose $1B month over month; last year’s $9B essentially doubled. Jin estimates roughly $50B in revenue this year, with 90-95% gross margins. Its order book (likely) has exceeded $110B, and Jin thinks it could close $150B in financing. Monthly results plus a completed financing would create a “very, very large” shock to the short-OpenAI-chain trade.
• Gemini’s organizational structure is its Achilles’ heel. DeepMind is based in London and remains highly lab-oriented, focused on building a better general-purpose foundation model. “It is not good at commercializing products into B2B businesses, and it has no interest in doing so. Put bluntly, Google has no interest in this business.” Jin is relatively bullish on Grok: its ability to build compute clusters is strong, it does not need Azure, and its costs are the lowest, while Elon brings product strength. “The only problem is that it cannot get enough cards.” Meta and Microsoft will also join the model war, because if they do not, the cloud rentier model will be transformed.

5. Software turns from drug company to factory, and Nvidia is the “land god”

• The old software model is a drug company: once a patent is completed, it can be copied and sold at almost no cost, with seat fees and subscriptions layered on top. That is why the market assigns 20-30x, even 40x, EB multiples. The future is a factory model that sells tokens. Every interaction carries a cost—Jin’s chess program used roughly $10 of tokens—and “the more factories you have, the more tokens you can sell.” Supply is structurally insufficient. Who builds the factories? “Nvidia. Oh, there is no Nvidia, and it is over.”
• Nvidia’s moat was bought with tuition: “Every pit Nvidia steps into today becomes a barrier to entry tomorrow.” In 2024, it solved liquid-cooling and connectivity issues for 72-GPU large systems, with Hon Hai (likely) and Dell helping debug the full stack. When AMD tries to build the second system, “no one is willing to invest money to work through it with you.” Nvidia does not sell GPU cards; it sells large-system solutions. That is why HBM has been bought out: Hynix, Samsung, and MU prioritize capacity for the most reliable buyer. “Nvidia locked up capacity, leaving everyone else unable even to obtain capacity,” much as Apple occupied the best suppliers when smartphones took off in 2008. There is currently a shortage of 3M cards, while datacenters to be completed this year need another 4M. As TSMC improves CoWoS yields, Nvidia simply sells more dies.
• Even if Oracle’s CDS does not work, it still has to issue stock and sell businesses to buy cards because this is a Monopoly game: “The cards I buy now are land. When you land on the Taipei square, if you do not buy the land, you are finished when you land there in the future.”

6. Models ultimately eat people, not SaaS

• Jin rejects the instinct that models will eat SaaS. Using Cowork to program inside Excel is unlikely to eliminate Excel, and seats may not be cut quickly. “SaaS is actually performing very well.” Replacement must be viewed through 4 channels: replacing people, replacing software, replacing competitors—“Goldman uses it, Morgan does not, so Morgan’s business comes to me”—and improving efficiency. “Everything else is indirect. What actually gets killed is people’s work.”
• The scale is enormous: AI could replace 30% of the world’s 700M white-collar workers earning an average $25,000. “New things always have to kill the old before they can create new productive forces.” Could the economy reach outright deflation, like 1848, with excess productive capacity and no buyers? “That is possible, but definitely not now. Now is a deflationary boom.” Even current AI beneficiaries such as Palantir, Databricks, and Snowflake may not be the final winners: Claude and OpenAI produce the answers directly, the buyers of cloud capacity become them, and the gross-margin economics are being rewritten.