Summary

The year’s main theme remains U.S. outperformance versus China and Europe amid decoupling, but Trump’s policies have turned a one-way trade into two-way tail risk. The U.S. economy remains relatively stable and its assets relatively strong, but tariffs, immigration and DOGE could “put a hole” in the market; China and Europe could surge after excessively bearish pricing. The speaker therefore refuses to go all-in, preferring to “wait for the right setup”; he still sees the S&P 500 around 6,000–6,100 in Q1, but expects violent swings along the way.
The tariff story is not over; the pause simply moves targeted tariffs into the negotiating phase. The 25% tariffs on Mexico and Canada respectively involve illegal immigration, China re-exports and a 3% digital service tax; broader trade-rebalancing measures could emerge from late March to April, with another 10% on China, 10%–20% overall, and potentially 25% on European autos. “Relief now does not mean the impact is over.”
Hong Kong stocks and China ADRs have upside tail risk, but still look more like a dead-cat bounce within a downtrend. China’s sustained money printing since September, CPI’s return to 0.5 and the initial 10% tariff on China all helped repair expectations; demographics, property, debt, income distribution and overcapacity have not reversed. Short-term participation is possible, but long-term overweight positions are unwise—especially after Hong Kong’s rebound from roughly 18,000 to 23,000 and back to around 21,000, which has materially narrowed the margin of safety.
U.S. Treasuries are in a relief window, but April could bring another wave of dollar, rate and trade-war volatility. The 10-year yield has fallen from 4.8% to 4.5%, while the speaker still sees 4.5%–4.7% as the normal range; markets price only 1–2 rate cuts over the next year and almost none over the 2- to 3-year horizon. His specific trade is to buy a 2-month dollar straddle because “realized is already higher than implied.”
Amazon’s $100B, Google’s $75B and Meta’s $60B–$65B spending plans show that cloud providers will keep investing this year, but Data Center is no longer a “no-brainer investment.” Nvidia’s full-year demand visibility remains high; GB200 issues can be absorbed through 8-chip systems or a shift to GB300. But the application layer has yet to truly break out, while the Magnificent Seven’s earnings-growth advantage over the rest of the S&P 500 has narrowed from 30 percentage points to 7 and could be just 3 in 2026.
DeepSeek lowers the floor for model-use costs, not the compute demand behind frontier training. The speaker sees it as reverse-engineering GPT outputs and optimizing the path; OpenAI’s 97% cut to o3-mini pricing, down to 3% of the original, will instead encourage more companies to try AI. “This is mass adoption on an industrial-revolution scale,” but whether the eventual winner is TEM, another healthcare application or an unknown company remains impossible to call.
Nvidia’s real medium-term risks are ASIC share loss, slower growth and margin compression; the new opportunities may lie in SaaS, cloud, connectivity and a new defense-industrial system. If gross margin falls from 75% to 70% or even 65%, growth could settle at 15%–20% or 20%–30%, leaving the stock valued at 20x rather than 30x P/E. Meanwhile, if DOGE redirects traditional defense procurement toward drones, robotics, command systems and software, a new defense-industrial complex around companies like Palantir could reach “10x-scale”; whether it can cut U.S. borrowing by $2T or even several trillion dollars remains a huge question mark.

Deep dive

1. Two-way tail risks replace the one-way trade of the past 2 years

The speaker’s core view is unchanged: the U.S. still has the strongest relative economic and asset trends versus China and Europe. Japan is a special case because of its own inflation and potential exchange-rate demands. What has changed is the “extreme randomness, volatility and suddenness” of Trump’s policies, making it impossible to equate U.S. strength with simply being fully invested in U.S. equities.
U.S. downside tail risks come from immigration, tariffs, DOGE, Europe and excessive market expectations. But his view is that if a policy shock causes a selloff, “the overall trend is still good; it will come back,” making the hole in the market look more like a buying opportunity.
China and Europe face the opposite setup: weak underlying trends, but assets that could suddenly surge on lower-than-expected tariffs, fiscal stimulus or negotiated outcomes. The speaker summarizes the year’s strategy as “waiting for the right setup” and maintains his Q1 forecast for the S&P 500 at 6,000–6,100: modest upside, with a choppy path.

2. The market’s focus has shifted from economic data to policy

The central question the year before last was whether AI could drive capex and technology earnings; last year it was whether the U.S. could contain inflation and sustain growth under high rates. Both have largely been answered. This year, “the market’s biggest focus is really policy,” with tariffs at the top of the list.
The latest payrolls report was soft, but the speaker believes fires, cold weather and changes in immigration magnified the noise. The trend remains roughly 180K jobs per month: a strong prior report followed by a weak one does not mean the direction has changed. The unemployment rate was 4.0%, better than the 4.1% expected and still relatively stable.
His baseline case is gradually controlled inflation with continued economic growth, and this week’s CPI could also come in near expectations. What could actually change the macro path is not one month of data, but whether tariffs or DOGE create additional inflation, mass layoffs or even a recession.

3. Trump is pursuing both negotiated tariffs and systemic trade rebalancing

The first category is targeted tariffs: wield the “big stick” to create panic, then trade a pause for concessions. The 25% tariffs on Canada and Mexico initially sent crypto lower before U.S. stocks followed; even the U.S. Postal Service’s suspension of Chinese packages, followed by a resumption 2 days later, showed the level of policy noise.
Mexico faces demands beyond illegal immigration, including cutting off the China re-export trade and overseas manufacturing routes that developed after 2019. Canada could face pressure over its 3% digital service tax. The speaker’s summary is “kill the chicken to scare the monkey”: force the largest U.S. trading partners to concede before turning to Europe and other regions.
The joint involvement of Michael Rubio, the Treasury Department and the Commerce Department shows that negotiations have bundled tariffs, diplomatic alignment, relations with China, technology-company access and ideology into a single package. The U.S.-Canada-Mexico trade agreement may not be completed until 2026, and repeated pauses and restarts in the meantime would not be surprising.
The second category is broad tariffs targeting current-account imbalances, layered with reciprocal tariffs. The speaker estimates that the policy could enter active discussion from late March to April; tariffs on China could rise another 10%, the overall level could reach 10%–20%, and European autos could face 25%, none of which is fully priced by the market.

4. China’s rebound has catalysts, but no structural reversal

The long-term constraints he lists include demographics, the property-bubble collapse, debt and overcapacity. A nearly 20% annual decline in marriages means the demographic-deflation factor operates on a “decade”—a 10-year horizon—not something that can be reversed in 1 or 2 quarters.
His distribution picture is that roughly RMB140T of deposits are concentrated among the richest 1%–2%, while roughly RMB170T–RMB180T of loans are spread more evenly across 400M–700M urban residents. The private sector accounts for roughly 42% of income versus about 80% in the U.S. His conclusion is that systemic debt is too high and household and private-sector income is too low; these problems cannot be solved.
The rebound logic is also real: China has been printing money continuously since September, and CPI has returned to 0.5. Even if the absolute figure is debatable, the margin is better than in prior months; the first tariff round on China was also only 10%, below more bearish expectations. But even Xiaomi, one of the few quality Hong Kong-listed companies, still operates amid excess capacity in phones and autos. Hong Kong stocks rose from roughly 18,000 to 23,000 and now sit around 21,000, making valuations less attractive.

5. Falling rates offer temporary relief; the dollar-volatility window is in April

When the 10-year Treasury yield rose to 4.8%, the speaker already thought the market had gone too far and was unlikely to price in Fed hikes. Treasury supply in February and March is lower than in January, bringing yields back to 4.5%; he still considers 4.5%–4.7% the normal range.
The curve implies only 1–2 rate cuts over the next year, with rates around 4%, and almost no cuts over the 2- to 3-year horizon. Europe and the U.K., by contrast, have clear easing expectations. The U.S. is therefore “sailing with the wind,” while the euro and renminbi are rowing upstream; the dollar could strengthen again after the relief rally.
His executable trade is to buy a 2-month dollar straddle for the potential European trade war in March and April: current realized volatility is already above implied volatility. Treasury supply will also rise in April, potentially repeating January’s stronger dollar and higher rates. Liquidity in Bitcoin and other liquid assets should improve somewhat, but conditions “are not especially loose.”

6. Hyperscalers pledge to keep spending, but earnings realization is beginning to lag expectations

The common question facing the 4 major cloud providers is how long they can maintain their investment resolve as capex keeps rising and margins remain under pressure. This quarter’s answer is still firm, with disclosed figures of $100B from Amazon, $75B from Google and $60B–$65B from Meta.
The speaker estimates that capex can still grow roughly 25%–30% this year, while data-center infrastructure built earlier is beginning to enter the chip-installation phase, with most incremental spending going to chips. On his stated figures, Nvidia’s results could rise from roughly 120B to 200B, an increase of more than 50%, leaving full-year visibility very high.
GB200 still has small issues involving its 72 chips, connectivity and debugging, but the system can be converted to an 8-chip configuration, and orders can shift to GB300, which may begin shipping at year-end. He thinks Nvidia could earn $5 per share this year. He may hold only through this year, originally hoping to exit once the Q4 and Q1 issues are resolved and the stock reaches $180–$200.
The risk is the expectation gap. The Magnificent Seven’s earnings-growth advantage over the rest of the S&P 500 was roughly 30 percentage points a year ago, narrowed to 7 this quarter and could be just 3 in 2026. If the AI application layer still fails to truly break out, management will have to explain every quarter what returns justify “continuing to do capex.”

7. DeepSeek cuts prices and expands the potential user base

The speaker does not view DeepSeek as a substitute for training at massive scale. His interpretation is that it uses GPT outputs for a form of reverse-engineering-style learning, then optimizes a more circuitous problem-solving path. This is an open-source model challenging the cost structure of closed-source models, not mastery of frontier training.
His sharper formulation is that DeepSeek “has not even found the training gate”; training a massive cluster of models with 10T-scale parameters still requires Nvidia. OpenAI’s path is to train better models faster, then cut prices immediately at launch, depriving imitators of both time and pricing windows.
The 97% price cut for o3-mini, bringing it down to 3% of the original price, is a direct example of the cost curve shifting lower. Lower prices will bring in companies that previously found AI too expensive, so overall compute demand remains intact and could even enter “mass adoption on an industrial-revolution scale.”
The application winners remain unknown. The speaker cites TEM’s example of “doctors using GPT”: healthcare data and medical businesses are promising, but no breakout is guaranteed. Just as cheaper website development did not reveal in advance whether the eventual winner would be the Yellow Pages, Yahoo or Google.

8. Parallel compute will expand with certainty, but GPU excess profits may not last

The long-term technology view is that serial CPU compute is shifting toward parallel GPU compute, as the physical limits of Moore’s law force the industry to adopt new development environments and instruction sets. CUDA and Nvidia captured the first major wave of gains, playing a role similar to the early full x86 architecture.
As parallel computing matures, the industry could move toward leaner architectures resembling ARM. Google has been developing TPU for years, and Amazon is building Trainium, but the speaker is less optimistic about Trainium, believing that making it this year and ramping production next year may not go smoothly. Common algorithms such as ad recommendation can also be hard-wired into cheaper ASICs. Thus, even if Nvidia still grows 20% next year, some of its market share will inevitably go to ASICs.
If growth slows to 15%–20%, or only 20%–30%, and gross margin falls from 75% to 70% or even 65%, Nvidia may command only 20x rather than 30x P/E. “It remains unbeatable in model training,” but being unbeatable does not make any price reasonable.
Higher-margin investment opportunities therefore shift to both ends of the stack. To adopt AI, enterprises must put data in the cloud and call SaaS; inside the data center, ASICs, Arista and networking connectivity offer more incremental elasticity than GPUs. Data Center remains the main theme, but it is “not a no-brainer business.”

9. DOGE could hurt growth or rewrite the defense-procurement chain

The speaker noted that a team consisting of “one college graduate and 6 high-school students” had shut down “USA,” with the next target reportedly the Department of Defense. This phase is not yet enough to drag on the U.S. economy, but how many people DOGE ultimately lays off and how much spending it cuts must be watched closely.
Foreign aid sent overseas does not count toward the relevant GDP, while Department of Defense spending certainly creates spending. The key question is therefore not the budget slogan, but whether DOGE genuinely cuts expenditure or simply reallocates the traditional defense-industrial complex’s share from Raytheon and Lockheed Martin to a new defense system centered on companies like Palantir.
His analogy is that Xiaomi could produce a car in a few months, while BMW takes 3 years from design freeze to production; Apple displaced Nokia through its operating system, and the Russia-Ukraine war likewise demonstrated a “software victory.” If drones, robots, command vehicles and control software partly replace tanks and artillery, the resulting new defense orders could be “10x-scale.”
The speaker does not believe Trump can easily reduce U.S. borrowing by $2T or even several trillion dollars. Without a large-scale reduction in borrowing, the direction of the economic fundamentals has not changed. DOGE can improve efficiency, but it cannot cut too deeply; that remains to be seen. Even if tariffs in full add roughly 0.5 percentage points to PCE, inflation swaps have risen about 0.6 percentage points since Trump took office, so much of that is already priced. The real tail risk is a DOGE-induced recession. With events such as Germany’s election also in play, the final strategy remains to clear out Hong Kong and China rebound positions, wait for a policy-driven U.S. selloff, and “capture this dip.”

91. A Paper-by-Paper Walkthrough of DeepSeek's 9 Key Papers and Their Innovations — “The Game of the Brave”The AI Architect: Bret Taylor