Market Overview, March 3, 2026

Summary

• Short-term bearish, long-term bullish is the backbone of the episode. 孙浩 warned in January that positioning and the setup were unfavorable, that “the only way is down,” and that the market needed a correction before making new highs—but February “somehow didn’t go down,” contrary to his expectations, because valuations are already low (Microsoft and Amazon are back at last April’s levels), retail buying enthusiasm is “enormous,” and $100B in tax refunds landing in March and April provides a floor.
• February “was like a long year.” The market began to realize that AI is not merely sweet capex but a force dismantling legacy industries; fear spread from individual stocks to the software index and then to credit markets. The widening IG CDS basis is his biggest short-term concern: Oracle shows that once credit spreads rise, financing further capex becomes harder and the company may need to issue stock.
• Positioning determines the short-term path. Put skew is extremely wide, with investors buying downside protection in both equities and credit; banks are negative gamma at spot, while a 7,000-point gamma wall caps the upside. Institutions sold $6.7B in February and systematic strategies (CTA/vol control) sold $26B; elevated volatility is keeping leveraged money out, creating a new normal of weak rallies followed by fast 3-5% selloffs that retail investors buy.
• SaaS and To C have been sentenced to death. The premise for giving software a 30x EV is 15-20% future growth, but “even if your results are excellent…you still can’t prove that OpenAI or Anthropic won’t replace you.” Any business model with people on one side and content on the other—Ctrip, Meituan’s front end, Baidu, online insurance—will be “swept out,” while models with people on both sides, such as WeChat and Meta, survive.
• Silicon Valley’s line is that “AI is a Little Red Book: give one to anyone and they can revolt.” A 3-5 person team calling open-source model APIs can outperform Copilot, and “all the moats and barriers built by TMT companies over the past 20 years can be knocked down in an instant.”
• The fear points to the opportunity: there is no fundamental problem anywhere on the path of supplying compute. AI has already passed its singularity point this year; OpenAI and Anthropic could each add about $1B in B2B revenue per month and together potentially exceed $50B this year. Older-card rental rates have risen to about 2.45, with purchased cards paying back in 18 months. Nvidia’s strong earnings without a stock rally reflect hyperscaler capex reaching $660B, or 92% of cash flow, but he expects that concern to be disproven from Q2 onward: half of capex is construction with a 20-year useful life, and Nvidia earnings can still grow 50% next year.
• The war previewed a restructuring of defense orders. On day one of the US-Iran conflict, existing intelligence combined with Palantir technology “took out all of the other side’s leaders,” with AI accounting for more than 50% of warfare’s impact. Palantir received less than 1% of the $600B in direct Defense Department spending, leaving substantial room for AI, software, and drones to take share.
• The CPU rethink is the episode’s clearest portfolio-rotation signal. As AI workloads multiply, the CPU has become an accessory to Nvidia’s rack: “the GPU is the factory worker; the CPU is the shop-floor manager.” Grace through Vera use Arm, but many customers want x86; he believes the probability of Nvidia and Intel jointly launching an x86 CPU card is high, and it could happen soon. Intel, whose position he had previously cut substantially, sits on the right side of the narrative and will be watched closely at GTC.

Deep dive

1. Positioning is the short-term ceiling: a 7,000-point gamma wall caps the upside, while rallies lack force and selloffs move fast

• 孙浩’s January view remains intact: he is “very bullish” on AI over the long term, but short-term positioning and the setup are poor; the market needs a correction before it can make new highs. February “kept trying to correct but couldn’t,” and the fact that it fell so little surprised him.
• The structure breaks down as follows: put skew has widened sharply, with investors rushing to buy downside protection in both equities and credit; “when uncertainty is high, you don’t have as much courage to sell puts.” Banks are negative gamma at spot, accelerating selling on declines, while 7,000 points above the market is a gamma wall: calls are heavily sold, making it difficult to break through.
• All 3 major sources of flow are selling: dealer hedging, another group of institutions that sold $6.7B in February, and systematic strategies—CTA and vol control—which sold $26B. Who is buying? Retail investors: “buying enthusiasm is enormous,” helped by Nasdaq valuations returning to last April’s levels and $100B in tax refunds arriving in March and April. “If this had happened in 2000…it would already have collapsed. Somehow, it didn’t.”
• There are only 2 ways out: put buyers either “lose enough money” while paying negative carry, or the market falls rapidly to a level where volatility spikes and sellers naturally return to the put market. Until then, the new normal is a lack of leveraged money on the upside, followed by rapid 3-5% declines that are caught by dip buyers.

2. AI’s negative shock has reached credit—this is what can truly affect the market

• The key shift in February was that the market moved from seeing “AI as sweet capex” to recognizing AI’s impact on existing industries. Fear spread from individual stocks to the software index and then into credit markets. “This thing hasn’t even started yet”: anything on the scale of the Industrial Revolution will inevitably be “one general winning fame while ten thousand bones lie exposed.”
• Widening IG CDS is his biggest short-term concern: investment-grade credit basis is rising across the board, signaling that the private sector’s cost of funding is increasing. Oracle is the case study—once its credit spread rises, financing continued capex becomes more difficult, and “it may have to issue stock,” which could weigh on the share price.

3. Narrative and earnings have diverged: SaaS and To C “should all be swept out”

• The logic is straightforward: the premise for valuing software at 30x EV is 15-20% annual growth, but AI may disprove that growth. “Even if your results are excellent, they will think Anthropic and OpenAI can replace you in the future.” No one can prove they will not be replaced 5 years from now. Hedge funds are selling software companies “with enormous enthusiasm” through their prime books.
• The verdict on To C is equally blunt: any model with people on one side and a content provider on the other is being written off. “Ctrip, you have no value”; Meituan’s front end will be replaced, while the content library Baidu spent years building is worthless in the face of models. Consumer-to-consumer businesses with people on both sides, such as WeChat and Meta, are spared. The same applies to online insurance and financial-product distribution, where barriers to entry will become very low.
• Silicon Valley’s sharpest metaphor is that “AI is a Little Red Book: give one to anyone and they can revolt.” A 3-5 person team calling APIs can “definitely outperform your Copilot,” and “all the moats and barriers built by TMT companies over the past 20 years can be knocked down in an instant.”

4. The compute chain is relatively sound, and concerns about Nvidia will be disproven

• The front-end battle is wide open, and “we still don’t know who will emerge as the winner,” but every model ultimately needs compute. AI has passed its singularity point this year; OpenAI and Anthropic could each add about $1B in B2B revenue per month and together potentially reach more than $50B, “and this is all just the beginning.” Rental rates for older cards are rising to about 2.45, purchased cards pay back in 18 months, and B-series cards are rented out immediately once deployed.
• Nvidia’s strong earnings without a stock rally come down to hyperscaler capex reaching $660B this year, or 92% of cash flow. Shareholder returns have fallen from 43 to 16, prompting fears that hyperscalers will lack the cash to invest next year. His path to disproving that concern is that half of investment over the past 2 years is construction with a 20-year useful life, while the cards are the cash-flow assets. The concern should gradually be disproven by Q2; even if capex growth slows, “Nvidia earnings still have to increase 50%.”
• His rough breakdown of the compute value chain is that about half of compute gains come from software. Of the remainder, 60% goes to Nvidia/TPU integration, half of what is left to CoWoS advanced packaging, and only then to 3nm/2nm wafer fabrication. The last 2 steps are both TSMC, alongside HBM, optical modules, and CPUs.
• CPU is the position he is actively reassessing. As AI workloads multiply, the CPU becomes an accessory to Nvidia’s rack: “the GPU is the factory worker; the CPU is the shop-floor manager”—more work means a busier shop-floor manager. Grace through Vera use Arm, but many customers want x86. He believes Nvidia and Intel jointly launching an x86 CPU card has a high probability and “could happen soon.” He had previously cut Intel substantially; earnings may remain difficult this year, with recovery not arriving until 2026-2028, but Intel sits on the right side of the narrative—unlike SaaS, which has earnings but no narrative—and he will watch GTC closely.

5. The macro backdrop is not poor: war and tariffs are unlikely to drive inflation, while liquidity is improving

• On the first day of the US-Iran conflict, “existing intelligence combined with Palantir technology…took out all of the other side’s leaders.” Beyond the more than 1,400 missiles, AI accounted for more than 50% of the war’s impact. “AI can deter war, because if you start a war, the leader goes first.” Palantir received less than 1% of the Defense Department’s $600B in direct spending, implying that orders will shift toward AI, software, and drone companies. Investors should stop putting everything into Lockheed Martin.
• His judgment on the motive for the war carries an explicit hedge: he does not think it was about the petrodollar hegemony cited in China; instead, he thinks Trump’s domestic tensions may have been an important factor, including protests over the “is” issue, the Epstein files, and record-low approval. “I don’t know—it’s just that I think domestic tensions were also a very important reason.”
• The Supreme Court overturned the tariffs, but the substance remains largely unchanged: the Court confirmed that tariffs fall under the executive branch’s authority, after which Trump used Section 122 to add 15% and conduct a 150-180 day review of other countries’ tariff rates. “Compared with the effective tariff that was canceled, there isn’t actually much change.”
• The closing read on inflation and liquidity is constructive. War is unlikely to create inflation in the US; the real risk would be a wage spiral in services, and “we don’t see that.” Oil has barely reacted, while natural gas matters far more to Europe than to the US—“this war is throwing a brick into someone else’s outhouse.” Financial deregulation is also adding liquidity: the new ESLR can take effect on April 1; Goldman estimates $5T in additional lending capacity, equivalent to roughly $200B of capital injection, while JPM estimates about $2.8T. With the FCI continuing to improve and $317B in buybacks supporting the market, “the fundamentals of the US economy are very strong.”