Market Overview January 6, 2026

Summary

• Macro throughline: decoupling again after year-end 2026. The speaker’s framework is 2023-24 decoupling (growth only in the US, and only in AI), 2025 recoupling (capital flowed back out of the US, lifting EM, Europe, Japan and China), and renewed capital inflows into the dollar and US risk assets in 2026; with almost no expectation gap left in China, Japan or Europe and expectations deeply negative for the US, the US has substantial upside from a reversal in sentiment.
• Dollar view: it is now in a bottoming range. The dollar is no longer entirely a “petrodollar” but increasingly an “US equity dollar,” tied to the stock market; the S&P held at 22-23x from early 2025 to early 2026, with the gain driven purely by earnings. Rate cuts are already priced in, and at least some investors who have priced in the loss of US hegemony will return in 2026.
• AI valuations broadly have yet to show a major bubble because the shovel sellers are priced for the most conservative case. With this year’s capex expected at $535B and Nvidia generating roughly 9.2-9.5 EPS based on TSMC production throughput, a $190 share price implies roughly 20x earnings—“the shovel sellers are all being priced off your most conservative estimate of how many shovels can be sold next year.” A real bubble would have to wait for the third stage: OpenAI listing at a $5T or $7T valuation, or “some law-firm application AI getting bid into the stratosphere.”
• Buying memory is less attractive than waiting for CPUs. Memory and optical modules are “relatively easy to expand” and could become commodities in 2-3 years; “they may make insane money for the next two years and then lose money later.” NAND is being priced as if demand will become 10x HBM demand, while future capacity from China-based ChangXin and others may mean products that can be made domestically are not actually scarce; market mouthpieces can use the capacity narrative to trigger a deep correction. The real shortage is in the latest CPUs: DDR5X servers have 12 channels, potentially 8 cores with 12 sticks per core (96 sticks total, 2T each), which older CPUs cannot support; GPU+HBM cannot simply replace CPUs because of thermal and other constraints.
• TSMC is the gatekeeper for the entire AI complex. Aside from Intel, which can also fab CPUs, TSMC is the only company that can mass-produce both CPUs and GPUs; capex is expanding only about 40%, “backing into capacity expansion from massive demand and keeping semiconductors in perpetual shortage.” GPUs were short 3M units last year and 4M this year; even after CoWoS ramps in 2027, they will still be scarce. That is the fundamental source of certainty in Nvidia’s earnings.
• Cloud chips are replacing handset chips. “Qualcomm is finished.” An NV720 system costing more than $3M can absorb price increases; handsets cannot. Handset iteration is slowing, and Chinese phone makers have only 5% margins. MU is cutting consumer-electronics capacity—“I’m not doing your business”—to focus on data-center DDR, where prices are rising 60-70%.
• BTC is not a short. Large crypto investors around the speaker have not been selling; “all the big players have actually been buying and staying bullish since just above $80,000.” A key reason for this selloff may be BTC’s removal from the MSCI indexes; once the removal is confirmed in February, the overhang may be fully priced.

Deep dive

1. Macro framework: decouple → recouple → decouple again

• The speaker’s cycle narrative: 2023-24 was decoupling—after the pandemic, global growth was driven only by the US, and only by AI. 2025 was recoupling, with capital flowing back out of the US, intensified by Trump’s tariffs, lifting EM, Europe, Japan and China as every region played catch-up. Europe launched a large-scale capex cycle from late 2024, while China “directly more than doubled base money… printing money at an insane pace.” The world entered a global fiscal supercycle. But “real growth and tech growth are still in the US,” so after year-end 2026, decoupling is coming again.
• The dollar’s redefinition is the linchpin of this view: it is “no longer entirely a petrodollar; to a large extent it is now an US equity dollar.” US equities are not simply overvalued; they “stayed firm on earnings even as dollar capital was flowing out.” The S&P traded at 22-23x from early 2025 to early 2026, with gains driven purely by earnings. The conclusion is hedged: the dollar is now in a bottoming range, but there is no call on how far it can rise while rate cuts are still underway. Rate cuts are already priced in, and at least some investors who have priced in the loss of US hegemony will return in 2026. Uncertainty should gradually fade around the midterms, and some of it is already priced in.
• Weekend geopolitical events were folded into the same logic: Maduro was captured, and “we don’t really know what kind of conspiratorial event happened in between, but at minimum it declared America’s absolute dominance in the Americas”—relevant to the FX view of capital returning. The bigger risk is not entirely an AI bubble but geopolitics: a divided-party midterm election and an unclear bargain behind Venezuela.

2. AI’s three expectation gaps: certainty hidden inside uncertainty

• First layer: the split between Wall Street and Silicon Valley. Wall Street’s framework is: “This year’s $535B capex is fine, and I can see how much money can be made, but 2027-28 capex can’t continue, so I price you low”—like the railroads: “You can bid up the steel mills first, but someone has to use the railroads; I don’t buy it.” The speaker calls that logic self-contradictory: “If you don’t believe Nvidia can stay in shortage indefinitely, how can you give the highest valuation to memory, which is easiest to expand? And what you’re pricing isn’t HBM—it’s NAND.”
• Second layer: AI boosting productivity elsewhere. The market has “just started to see the first signs, but hasn’t priced it clearly.” Analysts’ emails are basically written by AI; the speaker showed a PPT generated by Gemini directly from two hours of conference-call notes—“crooked in places, crooked in others; just think about what AI will be able to do later.” Among listed companies, only a handful of logistics examples such as CHRW (C.H. Robinson) have been recognized.
• The third layer is the least priced: AI-native companies taking over legacy industries. The team has looked at an AI law firm, likely Harvey; “a four- or five-person law firm using AI to take 10% of the business of US lawyers is perfectly normal.” OpenAI’s Mercury plan is spending more than $30M to hire financial professionals and has “already completed everything investment banks can do.” Among listed companies, only Palantir, through Department of Defense orders, and Tempus AI, through medical data, have broken out. The bubble test is whether things you cannot believe attract widespread belief and investment—just as people believed the internet stories in 2000. OpenAI reaching a $5T or $7T valuation would qualify. We are still nowhere near that point.

3. Memory is maxing out the expectation gap—buy memory, wait for CPUs

• The core warning on memory assets: optical modules and memory are “relatively easy to expand” and can become commodities in 2-3 years. “They may make insane money for the next two years and then lose money later.” NAND is now priced as if its demand will become 10x HBM demand, which is uncertain. Products that China-based ChangXin and others can make may not be scarce later; market mouthpieces can promote China’s future capacity, drive expectations for lower gross margins and trigger a deep correction. Last night’s optical-module selloff was an example.
• CPUs offer the highest-certainty shortage: GPU+HBM cannot simply replace CPUs because of thermal and other constraints, while the real shortage is in DDR5 and DDR5X server memory—think of DDR as HBM’s sidecar. The latest servers have 12 channels, potentially 8 cores with 12 sticks per core, or 96 sticks total at 2T each. “You simply cannot use an old CPU.” The speaker retains a self-hedge: “I don’t know whether this chart is drawn correctly; I don’t work directly in Double E.”

4. TSMC is the gatekeeper; GPU scarcity is the root of everything

• The industry’s first-principles logic is simple: “If GPUs are no longer scarce, what use are optical modules? GPU scarcity is the root of everything in AI.” Shortage can be measured by how many data centers have been built and how many cards they can hold: GPUs were short 3M units last year, 4M this year, and will remain in shortage even after CoWoS capacity ramps in 2027. That is the fundamental source of certainty in Nvidia’s earnings.
• TSMC is the “gatekeeper for AI babies.” Aside from Intel, which can also fab CPUs, TSMC is the only company that can mass-produce CPUs and GPUs. Its capex is expanding only about 40%, “backing into capacity expansion from massive demand” and keeping semiconductors in shortage. Its relationship with Nvidia is “extremely tight.”
• The long-term logic is cloud replacing edge: “Qualcomm is finished.” An NV720 system costs more than $3M, while the B-series chips themselves are relatively inexpensive and can absorb price increases; handsets cannot. Iteration is slowing—N2 chips may cost 2x N3 chips, without users needing the upgrade—and Chinese phone makers have only 5% margins, so they are all moving into cars. MU’s strategy is a signal: it is cutting consumer-electronics capacity—“I’m fucking not doing your business; automakers, fight for supply yourselves”—and focusing on data-center DDR, where prices are rising 60-70%.

5. 2026 macro basket and allocation conclusions

• The relatively optimistic feature of next year’s macro picture is that “employment will be fairly weak, but GDP will not be particularly weak.” With GDP holding up, corporate earnings should be good and the cost base should remain manageable; weak employment will also force the Fed and fiscal policy to become more aggressive. That is an important macro driver for risk assets.
• On allocation, the speaker reiterated that 2 assets look relatively safe this year, but investors should “buy the dip during a crash, not chase strength.”
• On BTC, after speaking with large crypto investors, “the big money around me still hasn’t sold.” Large players have been buying and staying bullish since just above $80,000. One important reason for the selloff may be BTC’s removal from the MSCI indexes; once the removal is confirmed in February, it should feel like “the bad news is out.” So “BTC is definitely not something you can short.”