Summary

The 2026 base case is not an AI bubble bursting, but elevated valuations grinding sideways as growth catches up. Bruce estimates QQQ at roughly 24.9x, toward the low end of the 25–27.4x elevated range, though still expensive in absolute terms; fiscal stimulus and a Fed focused on preserving growth can support fundamentals, but valuations will struggle to break decisively above the “hard ceiling” of the past 4 years. Returns will depend more on EPS growth than further multiple expansion.
What bursts the AI bubble is not the scale of financing, but the market losing faith in capability gains and commercial returns. Enterprises abandoning adoption, model capabilities plateauing, costs failing to fall, or ROI remaining structurally unattractive would be what drives hyperscalers to cut capex; for now, bank credit is growing at an annualized 5%–6%, with capital still flowing into data centers through shadow banking. As 泓君 summarized at the end of the episode: “Money is cheap; faith is scarce.”
ASICs will take share, but 2026 looks more like parallel expansion with GPUs than an immediate replacement of NVIDIA. GPUs retain a dominant position through general-purpose flexibility and cloud-platform neutrality, while ASICs lower costs for specific workloads, especially inference; every GPU and TPU instance is running at 100% utilization, and Hopper instances from several years ago still command high prices, showing that AI capacity remains undersupplied overall. “This is not a fight to the death”; the pie is still growing.
Google has moved from a low-valuation recovery trade to an EPS-delivery story, and a first-tier model is merely the entry ticket. Gemini moved the market’s valuation of Google from roughly 14x to 27–28x within months and created a premium of about 3–4 turns over Meta; the next leg higher must come from upward EPS revisions in search or cloud, while proving that AI Mode and chatbots will not seriously cannibalize existing search monetization. “A first-tier model is necessary, but not sufficient.”
OpenAI’s valuation depends on whether it ultimately becomes an iOS-style gateway or an infrastructure platform that is difficult to monetize. In the near term, it must prove that advertising and agent commerce can generate revenue; over the long term, general-purpose agents must enter white-collar salary budgets. Outside coding, however, adoption remains constrained by domain knowledge, vertical data and compliance. If value accrues to the “capability” of vertical agents rather than the underlying model’s “intelligence,” Bruce believes OpenAI could become a “value trap.”
The biggest tail risk may not be a gradual AI cooldown, but Europe or Japan triggering an abrupt tightening in financial conditions. Bruce assigns roughly 5%–10% probability to an AI slowdown scenario and expects it to build over 2–3 months; non-AI cyclical sectors could still fill the gap with fiscal and monetary support. The more plausible black swan is that aggressive European fiscal policy loses market support, or Japanese fiscal expansion collides with BOJ rate hikes driven by inflation, ultimately causing liquidity to contract and risk appetite to collapse.
Portfolio adjustments in 2026 should simultaneously target earnings delivery, factor mean reversion and confirmation of a non-AI recovery. In the fourth quarter of 2025, high-beta, so-called junk stocks and value significantly outperformed high-quality companies; Bruce links value’s strength to persistently rising real rates. The NVIDIA/GPU narrative has also split toward Google/ASICs, while memory and optical communications are growing faster. 任阳 added that although he had remained firmly bullish on compute and capex since the DeepSeek moment, he still missed the areas with the sharpest marginal changes. Traditional financials and consumer stocks are already strengthening, while real estate remains largely a trade on affordability-policy expectations. Ultimately, investors need to “follow the data,” especially whether PMI, still below 50, can genuinely recover.

Deep dive

1. The Most Valuable Call of 2025 Was the Policy Turn; the Most Painful Miss Was Marginal Elasticity

Bruce turned cautious in mid-February as financial conditions grew too tight and the market had priced away too much of its growth potential, allowing him to avoid part of the pullback that followed through late March. The tariff shock in April was a surprise—“we got hit too”—but his real concern was that Trump’s challenge to legislative independence could compress valuations.
The signal to add exposure came when Trump reversed course rapidly over a single weekend. Bruce judged that the economy would find a new equilibrium and that tariffs would not permanently impair its growth potential. By August 17, with the Fed actively choosing to support growth, he had shifted from caution to the view that investors “should not be overly pessimistic.”
The framework after that policy turn was for valuations to remain range-bound at elevated levels of 25–27.4x while growth caught up. By Bruce’s measure, year-on-year growth associated with QQQ rose from roughly 9% midyear to about 24%, resembling the combination seen from 2017 through the third quarter of 2018: elevated valuations with growth catching up.
The real miss was the fourth-quarter divergence: “high-quality companies were actually no good,” while high-beta stocks, “the junk names” and value led instead. Bruce believes value’s strength is tracking persistently rising real rates. The industry narrative also shifted from NVIDIA/GPU to Google/ASICs, memory and optical communications, with the latter two showing stronger growth, earnings elasticity and multiple expansion. 任阳 added that despite remaining firmly bullish on compute and capex from the start of the year, including after the DeepSeek moment, he still missed the areas undergoing the sharpest marginal changes.

2. Narratives Ignite the Fire; Earnings and End Demand Keep It Burning

任阳 described Google as “a process of the market being educated.” He had initially believed search was more vulnerable to AI cannibalization than e-commerce, social media or office software, but did not expect Gemini to reverse consensus within months and push the valuation from a trough of roughly 14x to almost double that level.
任阳 rejected treating narrative and fundamentals as a binary choice: narratives usually begin with “a seed of fundamentals” and also follow the stock price. Ultimately, investors still have to return to valuation, earnings power and risk-reward, because consensus shifts extremely quickly in the early stages of AI.
Bruce offered SaaS as the counterexample. HubSpot, CRM and similar names promoted AI agents and rallied for a time from late 2024 into early 2025, only to fall steadily after February. They went from AI winners to AI losers, while non-AI end demand still failed to recover: “The narrative was bad, and the existing business could not get going either.” It was a double hit.

3. Limited Valuation Upside in 2026, with Fiscal Policy and Liquidity Supporting Growth

Bruce’s overarching framework is that “a country is a company”: fiscal and monetary policy affect not only valuation, but also the fundamentals. Even if the Fed retains a bias toward cutting rates, U.S. equities at roughly 25x are already near their limit, and he still struggles to see a genuine break above 27.4x.
The One Big Beautiful Bill Act primarily rewarded corporate capex in 2025, but in 2026 will redirect some support toward individuals through refunds of 2025 taxes and lower personal income taxes in 2026. The first-quarter total is roughly $90B, or about $2,000 per U.S. household, which could support the U.S. economy through the first half.
Non-AI consumption, real estate and traditional manufacturing were repeatedly priced in ahead of time over the past several years, but never truly recovered. Bruce believes looser fiscal policy and a Fed supporting growth will again create the conditions for a 2026 recovery, but the recovery still needs to be validated by data; it is not an established fact.
Bitcoin’s fourth-quarter weakness also has a liquidity explanation: “The Fed’s balance sheet is no longer big enough.” Funding markets at the foundation had become too tight, forcing leveraged assets to unwind. The December Fed meeting focused on addressing that problem; if the repair proceeds smoothly, leveraged assets will benefit, but if funding remains tight, the pressure will persist.

4. Google Has Moved from a Recovery Trade to an EPS-Delivery Test

任阳 believes frontier models have very short shelf lives: Gemini, Anthropic, OpenAI, xAI and even second-tier players could trade leadership every 3–6 months. The base assumption should not be that any one company has already achieved an unassailable lead.
Google now trades at roughly 27–28x and carries a premium of about 3–4 turns over Meta, the exact opposite of its previous discount. Valuation is already “pretty full,” so the next phase must be driven by further upward EPS revisions in search or cloud.
A first-tier model is merely a “necessary but not sufficient condition.” Without one, Google could lose its eligibility to compete for the AI information gateway; with one, it still has to rebuild search productization and monetization so that AI Mode and chatbots can change the shape of the interface without consuming the high-margin search business that already exists.

5. ASICs Will Take Share, but GPUs Remain the Neutral Foundation

任阳 noted that ASICs are not a new story: Google has been developing TPU for 10 years, Amazon has Trinium and Inferentia, and Microsoft and Meta are also developing their own chips. GPUs will remain highly dominant in the near term, while ASICs provide more efficient support for specific workloads.
Among leading models today, Gemini has the most complete vertical integration with Google’s chips and cloud platform. OpenAI, Anthropic and others remain relatively independent, while each cloud provider wants to support as many models as possible. GPUs are naturally suited to this industry structure because they are general-purpose and do not bind users to a single cloud platform.
If independent model companies fall out of the top tier or explicitly align with one cloud provider, choosing a model becomes equivalent to choosing a cloud platform, and the hardware-software coordination value of ASICs would expand sharply. But chip design runs on a 2–3-year cycle, requiring the industry to forecast algorithmic demand in advance; software changes midway can affect production planning and tape-out, making it a case where one move affects everything.
陈茜 suggested that TPU expansion could be constrained by upstream and downstream dependencies as well as Broadcom’s priorities. Bruce then emphasized that the competition is continuously evolving. NVIDIA has already absorbed ASIC design concepts such as Tensor Core and is using CPX to improve inference efficiency; with the acquisition of Grok, it will not sit still and wait for share to erode.

6. Neoclouds Take the Dirty Work, while AMD Is Caught Between the Second Source and In-House ASICs

任阳 believes AI capacity remains undersupplied: every GPU and TPU instance is running at 100% utilization, while Hopper instances from several years ago remain expensive and fully loaded. 2026 is therefore still a period when “the pie is getting bigger,” rather than one in which different forms of compute immediately cannibalize one another in a zero-sum fight.
AMD still has room to grow, but its strategic position is more awkward. It used to be NVIDIA GPU’s second source or backup supplier, but migrating from NVIDIA GPUs to AMD GPUs requires cloud providers and AMD to complete software adaptation together. With cloud providers’ limited engineering capacity now prioritized toward their own ASICs, AMD’s internal priority has fallen accordingly.
任阳 defines Neoclouds as doing the “dirty and exhausting work” for the major players: CoreWeave takes compute capacity wherever it can find power, while Oracle has shifted from a traditional value play to taking on financing, financial risk and inventory exposure. Their upside elasticity is high, but they also sit at the highest-risk end of the portfolio and are the names most likely to be “hunted down by the market” when the cycle ends.
Oracle and CoreWeave share prices can serve as a read-through on confidence in the AI buildout. Bruce believes investors might be less worried about OpenAI’s ecosystem—and Oracle might perform better—if OpenAI had not thrown its financing plan at the market aggressively all at once, but instead rolled it out in stages.

7. The AI Bubble Does Not Lack Money; It Lacks Evidence Capable of Breaking Faith

Bruce’s definition is direct: “A bubble does not burst by itself; it needs a reason.” A consensus to cut capex would form only if enterprises refused to use AI, model capabilities stopped improving, costs could no longer decline, or returns on investment became unattractive. For now, the major players’ investment confidence remains extremely strong.
Funding is not yet a constraint either: U.S. bank credit to the private sector is growing at roughly 5%–6% annually, with substantial capital entering data centers through shadow banks and institutions such as Blackstone. After lying dormant in 2023 and 2024, this expansion has only recently begun to accelerate.
Bruce compared it with shale oil: the U.S. burned capital from 2009 through 2014, and it ultimately took a Saudi price war to break the faith. He put the 2014 funding scale at $900B, compared with roughly $1.5T today. Internet-related investment also grew at 20% annually for 7–8 years, while the current AI cycle has only reached 20% for the first time.

8. OpenAI Must Prove It Is the Gateway, Not an Infrastructure Platform That Is Hard to Monetize

Asked whether OpenAI would be worth buying if it went public, 任阳’s answer remained, “Hard to say.” In the near term, he wants to see mature productization and monetization paths for advertising and agent commerce; over the long term, the question is whether general-purpose agents can enter the salary budgets of white-collar and knowledge workers and deliver on the “sea of stars.”
The difficulty is that enterprise applications contain substantial domain knowledge, vertical data and compliance requirements; a single general-purpose agent cannot solve everything. 任阳 is therefore optimistic about the direction over the medium and long term, but does not believe most industries will adopt it as quickly as coding has.
Coding agents lead because the data is well defined, programmers are eager adopters, and the workflow still contains substantial human feedback. If spending a few hundred extra dollars a month can double coding productivity, users naturally feel that it “makes sense.” Other industries do not yet have the same clear feedback loop.
Bruce offered two extremes. If OpenAI becomes an iOS- or Windows-style gateway that everyone must pass through and that is difficult to commoditize, it would be extremely valuable. If models converge and all the value accrues to vertical agents that call the models—“the model has intelligence, but what is truly valuable is capability”—OpenAI could become a value trap.

9. Apple Can Plug into Other Models, but It Cannot Lose the Next-Generation Gateway

Bruce is unsure whether Apple is waiting for models to mature before entering at a better price-performance point, or whether it genuinely cannot find a solution. Talent continues to flow to different AI companies, while neither on-device AI nor in-house frontier models has shown obvious progress. He is especially skeptical about “how much value an entirely on-device model can really deliver.”
任阳 has “not much expectation” for Apple’s own frontier model. Connecting to Gemini or ChatGPT is entirely feasible, just as Apple’s browser defaults to Google Search. The key issue is not who owns the model, but which side ultimately has greater bargaining power between models such as Gemini and Apple’s ecosystem.
The real risk is that AI creates new killer hardware and rewrites the gateway. 任阳 is not bullish on the AI Pin, and sees smart glasses or an unknown form factor as more likely. But every device currently imaginable would struggle to fully replace the smartphone; the more realistic role remains supplementation.

10. SpaceX’s National-Security Scarcity Outweighs Traditional Valuation

Asked how much space-based data centers could add to SpaceX’s valuation, Bruce said the question is not especially meaningful in isolation. The world is rebuilding redundant supply chains under the pressure of trade wars and other forces, making national security an investment driver. SpaceX and China’s Long March rockets both combine national-security relevance with scarcity, placing them outside ordinary valuation frameworks.
His market figures are roughly $810B for SpaceX today and an expected IPO valuation of $1.5T–$2T. His own approach might be to “put a little money” behind Elon Musk while holding a scarce asset, but he would not go all in.

11. The Real Black Swan Is More Likely to Travel Through Sovereign Fiscal Policy and Liquidity

Bruce treats models reaching a capability plateau, inadequate ROI and hyperscalers cutting capex as one causal chain, assigning a 5%–10% probability to an AI slowdown in 2026. It would build gradually over 2–3 months, while non-AI industries could fill the gap with policy support, so it would not look entirely like a sudden black swan.
More concerning would be aggressive European fiscal policy losing market support, or a new Japanese government expanding fiscal policy while the BOJ continues hiking rates because of inflation. Using the U.K. as an example, Bruce said that if markets stop supporting a government’s aggressive fiscal policy, the result could be a short-term liquidity contraction.
Bruce’s conclusion is that the real shock would come from a sudden tightening in financial conditions and a collapse in risk appetite, not from an AI data inflection point that can be observed continuously. This is not the base case, but it could become a scenario.
Beyond the popular names, he is more interested in Synopsys and Cadence: “They are like the blood in the semiconductor industry.” They are not sexy and lack a Huang-style figure constantly telling the story, but they carry substantial industry importance and technological barriers to entry.

12. Non-AI Rotation Is Already Being Anticipated, but the Data Has Not Confirmed It

After November and December, U.S. financials and consumer sectors strengthened relative to the market over the past 2 months, while the Dow Jones, with greater exposure to non-AI cyclicals, also began testing the upside. Ahead of the midterm elections, Trump began emphasizing affordability, calling for lower inflation, higher wages and more affordable housing, while previewing real-estate reforms in the first quarter.
Bruce’s reservation is: “You still cannot see it in the data; this may just be faith.” The market traded the same recovery in early 2025, only for it to fail to materialize. Whether 2026 is different should be determined by continuing to “follow the data.”
The conditions for recovery do exist: U.S. PMI remains below 50, while the world is rebuilding supply chains and infrastructure and adopting expansionary fiscal and monetary policies, leaving room for cyclicals to move higher. If PMI were already at 60, it would instead suggest that the cycle was nearing a top.
At the portfolio level, Bruce is watching for mean reversion between low-quality high-beta stocks and high-quality equities, as well as renewed convergence between the NVIDIA/GPU chain and the Google/ASIC chain. The market cycle is already in an elevated-valuation zone, with QQQ at roughly 24.9x and a ceiling of 27.4x. But the growth cycle still has room to run: “The AI train may even accelerate as it moves forward.”

Anthropic's $10BN Raise, a16z's $15BN Fund and ElevenLabs at $11BN New Year Livestream 2: Tesla FSD and the Commercial War in Autonomous Driving