Market Overview, December 9, 2025
Summary
- The Fed is highly likely to cut rates this Wednesday, but the core trade has shifted from “whether it cuts” to how far it can still go internally: the policy “train has left the station,” and whether the chair is Powell or Kevin Hassett, whom Trump may nominate, the tone will be soft. The speaker believes 2 cuts in 2026 are already more dovish than many investment banks; PCE is still at 2.8, and even if it falls back to 2.2–2.5 after tariff effects fade, room for cuts is limited. More liquidity will come from bank deregulation and various forms of balance-sheet expansion. There were already 2 dissenting votes against the last cut, while doves want 50 basis points; the split is the path variable.
- The market has priced a December 18 Japan hike at roughly 90%, but the speaker’s high-conviction call remains a weaker yen and stronger equities in 2026: a near-term hike will lift the JGB curve and compress the carry trade, but over the medium term it will be constrained by the 高市 government’s fiscal expansion and the BOJ’s structurally dovish stance. Services and labor costs are rising together, and inflation swaps are moving higher, increasing pressure on BOJ to hike; but it is “a bit like your mother hitting you… you did do something bad, but she really can’t bring herself to hit that hard,” and with limited independence, a weaker yen will ultimately continue to provide liquidity for global risky assets.
- The speaker expects the US 10-year Treasury yield to end the year around 4.2%; the earlier move to 4% was a “free-money trade” driven by reduced supply during the government shutdown, uncertainty, and safe-haven demand, and is difficult to sustain. He expects the drag from tariffs to fade in 2026, with fiscal expansion turning positive in Q1 or Q1–Q2 and feeding through to the economy in Q2–Q4; absent an extreme fiscal accident, the fundamentals remain strong.
- Short-term positioning also does not support a deep selloff: CTAs cut about $70B in November, and after last week’s unusually rapid drop in vol, CTAs and vol-control funds would need to buy back roughly $7B a day even if the index moves sideways. The probability of a pullback after Wednesday’s cut is therefore “quite high,” but the speaker does not expect a repeat of this April and thinks it may not even reach last August–September’s level; “everyone calling for a bearish market is long,” and any pullback is an opportunity.
- NVIDIA’s key edge is not that TPU is cheap, but its performance per watt, flexibility, and deliverable capacity when power and space are constrained; TPU in fact proves that the scaling law works, increasing total compute demand. The speaker believes ASICs and TPUs will struggle to replace GPUs while models continue to iterate rapidly, with TSMC capacity the real bottleneck. He expects 2026 EPS of about $9.5, growth accelerating from roughly 60% to 70%, and a stock price in the low $180s implying 18–19x forward P/E, excluding China-bound H200. “Putting all of this together, no chip can beat NVIDIA” is his strongest conclusion.
- The single point of failure in the Oracle short thesis is OpenAI’s ability to pay: Wall Street thinks ORCL is leveraging up to build compute centers against its orders, and if the customer stops paying, the entire chain ahead of it collapses; the speaker counters that OpenAI’s monthly maintenance cost is only $1.5 per MAU. His categorical view is that OpenAI will not go under, and even if it runs short of cash, it could be acquired by Apple, NVIDIA, or others, so the orders remain valid. On that basis, he believes the short frameworks around AI “left foot stepping on right foot” and Oracle buying from itself and selling to itself do not hold.
- The 2026 semiconductor theme can be summed up as “everything useful will be in shortage”: CPUs should rise at least 20% next year, roughly 20%–30%, and could even reach 30% in 2027; this makes Intel, with flexibility to shift capacity from Intel 3/4, a more direct beneficiary than AMD, which is constrained by TSMC. The speaker says 18A yield is about 70%, improving 2%–5% per month and settling at roughly 2% thereafter; internal tape-outs become cost-viable at 60% yield. Combined with a potentially rapid Apple investment, his valuation framework gives the CPU business at least RMB200B and Foundry roughly RMB200B at 2x P/B, or RMB400B in total with all other businesses valued at zero—and he believes Intel is nearing a turning point.
Deep dive
1. Rate cuts are a done deal; Fed divisions will determine the scope of 2026 easing
The speaker treats Wednesday’s rate cut as essentially certain. Even with little usable data, Friday’s PCE was only marginally better and still at 2.8; “the train has left the station,” and the market should focus on how much internal conflict is exposed after the meeting.
His 2026 base case is 2 cuts, already more dovish than many investment banks. Inflation could return to 2.2–2.5 as tariff effects fade, but rates cannot be cut indefinitely, so larger sources of liquidity will come from bank regulatory easing, balance-sheet expansion, and similar tools.
Kevin Hassett has a high probability, in his view, of being nominated by Trump. Whoever chairs the Fed, policy will ultimately remain soft, though economic facts still matter. There were already 2 dissenting votes against the last cut; doves want 50 basis points, while hawks want no cut, with some even more aggressive views. The degree of division will determine how much farther the Fed can go and how it gets there—not whether its overall direction remains dovish.
2. Japan hikes in the short term, but policy remains dovish over the medium term and the yen stays weak
Japan’s government bond auction was still well received last week, but the MOF is unlikely to materially reduce issuance. The 10-year and 30-year curves could steepen further, with the 30-year yield continuing higher. At the same time, services and labor costs are rising together and inflation swaps are moving higher, putting pricing for a December 18 hike at roughly 90%.
Japan has long been a source of global liquidity. A carry trade funded in low-yielding yen and reinvested in US Treasuries will contract in the short term as a hike shifts the curve higher. The speaker acknowledges the disruption but does not believe it will become a lasting reversal in liquidity.
The BOJ’s response to inflation and deflation is asymmetric. If the Fed’s sensitivity to employment were “multiplied several times,” that would approximate the BOJ’s attitude toward deflation risk. His analogy: “It’s a bit like your mother hitting you… you did do something bad, but she really can’t bring herself to hit that hard.”
The medium-term forces are the 高市 government’s fiscal expansion and reflationary policies, including support for defense, AI, nuclear, and shipbuilding. The BOJ is also not fully independent and will struggle to keep rates very high for an extended period. The speaker’s conclusion remains that “a weaker yen is a major trend,” leading to higher equity prices and more abundant global risk liquidity.
3. A 4.2% Treasury yield and system flows continue to support risky assets
The speaker has consistently expected the US 10-year yield to end the year around 4.2%. The earlier move to 4% came as the government shutdown reduced Treasury supply while uncertainty and safe-haven demand rose at the same time; under those conditions, it was “a free-money trade for macro markets,” and could not persist for long.
The growth logic is a crossing of 2 forces: the drag from tariffs gradually shrinks while the contribution from fiscal expansion grows. Based on the Goldman Sachs estimate he cited, the net impact could turn positive in Q1 2026 or between Q1 and Q2, then show up in economic data in Q2–Q4. Unless fiscal policy produces an extreme accident, the fundamentals remain favorable.
CTAs cut about $70B in November, and the systematic positioning represented by CTAs and vol-control funds is potential buying power. After last week’s rapid decline in vol, even a flat market could see roughly $7B flow in each day from Monday through Friday. The odds of a decline after the rate cut are therefore not low, but the speaker thinks it could at most approach last August–September’s conditions and might not even reach that level; a pullback would instead be an opportunity.
4. TPU expands compute demand without weakening NVIDIA’s moat
The speaker compares chips across 3 dimensions: performance per watt, flexibility, and cost. TPU is indeed cheaper, but a data center is like a factory with limited power and capacity. When supply is extremely tight, equipment cannot be judged solely by unit cost; the relevant question is output efficiency and flexibility against constrained resources.
What TPU really proves is that the scaling law still works: adding more compute produces better models, so total demand expands rather than contracts. Continued model iteration also requires flexibility. His view is therefore that ASICs and TPUs are still a long way from replacing NVIDIA, and may never do so.
In the supply game, hyperscalers are buying TPU to pressure NVIDIA and obtain more allocation directly, bypassing channels such as CoreWeave and Lambda. But the ultimate bottleneck remains TSMC’s tape-out capacity. The speaker also believes Google is one of TSMC’s largest customers: Google’s DeepMind continues to secure NVIDIA chips, while its Cloud division sells TPU externally and tests the outside ecosystem. He goes so far as to say that as long as TSMC has capacity and 张忠谋 is still there, NVIDIA will not face a genuine supply shortage.
The speaker expects NVIDIA’s EPS to reach $9.5 next year, with growth accelerating from roughly 60% this year to 70% and margins continuing to improve. At a stock price in the low $180s, that implies 18–19x next year’s P/E, without including H200 sales to China under the arrangement Trump has described. A high share price does not necessarily mean a high valuation: “you have to look at its earnings.”
5. Oracle shorts have concentrated all the risk in OpenAI’s ability to pay
Wall Street is shifting part of its AI allocation toward healthcare and banks while concentrating shorts in Oracle. The full thesis is that Oracle takes large orders and leverages up to build cloud and compute centers; if customers such as OpenAI cannot pay, the entire chain of orders, financing, and capex collapses.
The speaker’s counterexample is that OpenAI’s maintenance cost is only $1.5 per MAU per month, far below the spending intensity of Douyin, Xiaohongshu, and early internet platforms competing for users. OpenAI also owns the model. His categorical call is: “It will not go under”（它不会倒掉）.
Even if OpenAI ultimately runs short of money, he believes Apple, NVIDIA, or another company would acquire it directly, so the orders would remain valid. Oracle simply “has an order in hand, then goes to build a plant” to fill it. If the core customer does not disappear, high leverage does not automatically imply “left foot stepping on right foot” or self-buying and self-selling.
6. Agents push CPUs into an “everything useful will be in shortage” 2026
The speaker compares compute infrastructure to a highway. The market is still asking whether there will be cars once the highway opens, while the reality is that “all the cars are already stuck; there just isn’t a road.” He expects new supply to run at full capacity next year. Whether 2027 remains fully booked is unknown, but next year’s investment will push optical modules, memory, and every useful class of chip into shortage.
Sora is not the cause of the memory shortage; it merely reheated a shortage that already existed. The more important new variable is the agent. Traditional model calls rely mainly on GPUs with CPUs in support, while agents make continuous, programmatic calls to both CPUs and GPUs. The more use cases emerge, the greater the demand for CPUs.
The speaker estimates that CPUs will rise at least 20% next year, roughly 20%–30%, and “could easily reach 30%” in 2027. AMD still has a strong performance-per-watt offering, but its TSMC capacity lacks flexibility. Intel can use Intel 3 and Intel 4 to shift part of its low-margin consumer-chip capacity to servers, so its capacity may be the first to be fully claimed. His summary is that next year, anything useful in semiconductors will likely be in shortage.
7. Intel’s turning point comes from capacity, yields, and a valuation reset
Intel’s recent rerating cannot simply be credited to its new CEO 陈立武 (Lip-Bu Tan). The speaker attributes the foundation to Pat Gelsinger’s “5 tracks” over the past 3–4 years and the creation of Intel Foundry. Semiconductor results require a long incubation period; he had already judged Q3 2025 to be the hardest quarter, with conditions gradually improving afterward.
The original design and PDK for 18A were not intended for external customers; 18A-P was the version redirected toward external use, while 14A was designed for external customers from the start. The speaker says 18A benchmarks against N3 and approaches N3/N2 levels. The text also states that TSMC and Intel both had yields of roughly 70% at the time, while external-customer use would require about 90%. He says Intel’s yield was improving 2%–5% per month, later settling at roughly 2%.
For internally used wafers, Intel could reach cost parity with outsourcing at a 60% yield because the price paid to TSMC includes TSMC’s margin. 14A benchmarks against TSMC’s 1.6nm process. The speaker says Apple could tape out A19 in 2027, and possibly A20 as well.
The original Apple–Intel deal called for Apple to pay the R&D costs, but it is now more likely to become a direct investment. Based on his industry sources, the plan is largely worked out and investment is essentially settled as the structure; Apple could also provide the money relatively quickly.
Wall Street values Intel on the low-margin economics of its combined Foundry and product businesses, with some analysts putting the stock in the $20s. Semiconductor capital in Taiwan has begun chasing Intel. Using legacy foundries such as GlobalFoundries at roughly 1.7x P/B and TSMC at roughly 7–8x P/B as references, the speaker believes Intel Foundry is currently valued at close to 0x P/B. The CPU business is worth at least RMB200B and Foundry roughly RMB200B at 2x P/B, for RMB400B in total; even with every other asset valued at zero, Intel still looks cheap.
He also believes Lip-Bu Tan has the best management relationship with Wall Street among chip-company executives. At Intel’s last earnings call, the company set low expectations for future margins, gross margin, and earnings, which made the market more willing to accept the outcome. Taking together the 18A, 18A-P, 14A, and capacity logic, he believes Intel is nearing a turning point.