168X in Conversation with Rick: What Comes Next for AI After Anthropic’s IPO? #2

Summary

• Rick says Anthropic is “severely undervalued,” but won’t bet on its post-IPO share-price path. The host cites $4.6B in 2025 revenue and $16.2B in the first half of 2026, while GAAP losses remain close to $700M. Rick says convertible-debt-related liabilities explain much of the reported cost, while acknowledging he has not examined the debt line by line. His buy-in premise is an IPO valuation of “1.5,” with the unit unspecified. Asked whether the stock might follow SpaceX’s post-listing path of an initial jump followed by a decline, he says, “I don’t know, but I don’t care.”
• His core forecast is that leading model companies may shift from selling subscriptions to delivering outcomes within 6–12 months. If advances in math and physics translate into new materials, proteins and drugs, a $200 monthly plan will no longer be the main source of value; a new drug formula could go straight to Eli Lilly. The strongest models may also remain closed, making it harder for competitors to catch up.
• Even if monetization slows, Rick doesn’t think frontier-model companies can afford to slow compute investment. He sees a winner-take-all race in recursive self-improvement: “Going all in still may not be enough to win,” while holding back likely means losing. He reads Microsoft’s reported voluntary packages for 60,000 employees and six rounds of layoffs in two years as signs the company is cutting other costs to protect AIDC capex.
• He thinks AI borrowing will push “higher for longer” toward even higher rates, hitting real estate and overseas markets. He estimates AI-linked companies will issue about $300B in debt in 2026, adding financing demand equal to roughly 20% of U.S. government borrowing; that could rise to 30% in 2027. He has not worked through the figures in detail. He expects the 10-year Treasury yield to top 6% sooner or later and says 7% or 8% would not surprise him. If yields stay above 6% for long, he fears capital will flow back to the dollar and trigger another regional financial crisis.
• Rick has sold most of his Nvidia shares and shifted toward manufacturing, equipment and optical assets he sees as harder for models to displace quickly. He says CUDA’s moat is gone and worries that model companies designing their own chips will erode the value of fabless businesses. He is holding the rest of his Nvidia position mainly because he has owned the shares for less than a year and would face a higher tax bill if he sold. In his view, TSMC, Intel, ASML and some test-equipment makers benefit from the “time that can’t be shortened” in qualification and production.
• Memory remains his largest position. He would rather sit through a drawdown than miss a turn in the AI trade. He owns Micron, SK hynix and Samsung; he has left his shares untouched, though his options have lost considerable value. His thesis is that stronger models make each unit of compute more valuable, potentially lifting the value of older hardware too. “I don’t know when it’ll come back,” but he says investors should focus on “time in the market, not timing the market.”
• NeoCloud is the sector where he has lost the most; his key bet is that power contracts are undervalued. He separates existing fixed-rate debt, contracted revenue and pass-through power costs from the cost of future expansion. After counting queued power contracts as assets at his estimated value, an AI tool put the price-to-book ratio at about 1.6–1.8x. That adjusted figure depends on how the contracts are valued. He acknowledges their prices could fall and the company could still go bankrupt.
• Rick has sold all the Bitcoin he had held for years, yet still believes a long-term agent economy can only be built on crypto. He is stepping back because of uncertainty around AI-driven advances in math and quantum computing: he does not know how many logical qubits it would take to break elliptic-curve signatures or whether that is 5–7 years away; he thinks it could be just 2–3 years. Even if the risk never materializes, fear alone could trigger a steep selloff. He also plans to trim some Ethereum while retaining some crypto as a hedge against fiat risk.

Deep dive

1. Anthropic’s revenue surge and reported losses need to be considered separately

• The host opens with a sharp financial contrast: Anthropic generated $4.6B in revenue for all of 2025 and $16.2B in the first half of 2026, yet still posted a GAAP loss of nearly $700M. His question: growth is fast, but are the costs too large to ignore?
• Rick attributes much of the reported cost to convertible debt: as the equity value tied to early financing rises, the associated liability rises too. He recalls actual spending of about $10B and sees rare capital efficiency relative to current ARR, while qualifying the figure with “I think” and acknowledging he has not reviewed the debt composition line by line.
• He calls Anthropic “severely undervalued” and says he would buy if it IPOs at the “1.5” valuation he cites, without specifying a unit. His view comes from a quick read and AI-assisted analysis, not a completed, line-by-line review of the financials.

2. Model companies may move on to selling scientific breakthroughs

• Rick connects math, physics, materials and life sciences into a single value chain: stronger frontier math could unlock advances in physics, then new materials, proteins, cancer drugs, quantum computing and fusion. In that scenario, a $200 monthly subscription may soon be a poor measure of these companies’ potential.
• He recounts Sam Altman’s electricity analogy from a Stanford Q&A: electricity initially frightened people, and power companies sold households on light—they “sold light.” In the same way, most people may not understand tokens, but they can understand what a new drug or material delivers.
• The customer base could change too: a drug formula could go to Eli Lilly, a new material to Applied Materials, without chasing consumers one by one. His timeline is speculative: “maybe in another six months, a year at most,” leading companies will stop selling subscriptions and sell outcomes directly.
• As an example, Rick cites what he says was an OpenAI model used in late August or September to solve the Navier–Stokes equations, which has yet to be released to the market. He speculates that the strongest models will stay in-house. He explicitly calls the idea that leading companies may have an informal understanding about when to open access his own “very cynical” read.

3. Questions about monetization do not change his capex view

• The host’s challenge is worth preserving: could monetization in frontier research take much longer than subscriptions and APIs? Rick concedes it could. ARR may fall in the short term, perhaps even affecting capex; that is also why markets worry about upstream semiconductor demand.
• But he does not think ambitious frontier labs can voluntarily slow compute investment. Recursive self-improvement is a winner-take-all race: “Going all in still may not be enough to win.” Hold back, and you will probably lose.
• In his framework, CFOs must balance debt issuance, equity issuance, the share price and cost cuts while remaining “as aggressive as possible.” Slower near-term revenue realization does not remove the incentive to compete for the next level of intelligence.

4. He reads Microsoft’s repeated layoffs as a signal that it is protecting compute spending

• Rick cites reported voluntary packages at Microsoft for 60,000 people, along with six rounds of voluntary or involuntary layoffs over the past two years. He guesses that releasing the news ahead of earnings was Amy Hood’s way of telling the market the company would cut costs, cushioning concern about the coming capex figures.
• His comparison: one 20% layoff usually gives remaining staff more confidence than four rounds of 5% each. Microsoft’s repeated moves suggest that “plans can’t keep up with change,” including its original capex budget. He stresses this is a non-expert read; the host says cost cuts could help the share price but clearly hurt employees.

5. IPO fund flows and earnings growth are two separate questions

• The host asks whether Anthropic could follow SpaceX’s post-IPO path of rising about 50%–60%, then falling for months. Rick won’t bet on it: “I don’t know, but I don’t care.” If the stock drops, he will use the same money to buy more shares; if it does not, he accepts ending up with fewer.
• Could the “purest AI stock” pull money away from memory, optics and other hardware? Rick says it could. But Anthropic’s likeliest use of new funding is to buy more compute, which could still benefit the supply chain overall. The net flow shift is “hard to say.”
• He expects high rates to compress valuations and says the next leg up will depend mainly on earnings, both from model companies and businesses using AI to augment their operations. Bio is one area he has seen: he says related stocks have risen severalfold over the past two months. But revenue may grow faster than profit, and the market could still punish model companies.

6. He reframes open-source competition as a closed-loop race for compute

• Asked whether open-source models are 12–18 months behind, Rick does not take the premise at face value: “It’s hard to say whether the gap is 12 to 18 months.” He only thinks the gap could widen, contrasting closed models working on frontier science with open-source models still chasing scores on benchmarks such as SWE-bench.
• His first argument is “the model is the application”: models will eventually consume applications. The second is “compute is the model”: only sustained access to compute lets model companies accelerate recursive self-improvement. In his view, a model company without enough compute may have no investment value at all.
• The third is a conditional “a moment becomes eternity”: if model iteration, chip design, fabs and energy infrastructure form a closed loop, compute leads to better models, and better models bring in more compute. He thinks the feedback loop’s growth exponent is “probably greater than 3,” after which the leader may be nearly impossible to catch.
• He interprets Elon Musk’s moves as an attempt to overtake rivals before this full loop forms. Open-source models competing on token prices can certainly affect closed-model revenue; if the strongest models keep their APIs closed and sell outcomes directly, the fight for the same revenue pool looks different.

7. Both Chinese and U.S. AI stocks look oversold to him, but he has not studied Chinese names closely

• On China, the host mentions ByteDance, Tencent and Alibaba. Rick says he has only heard that Alibaba has secured compute and needs to make time to research it; he has not closely examined upstream or downstream A-share semiconductor names either.
• His view is conditional: as long as China has not given up on the AI race, it will keep investing all out, and battered supply-chain assets are worth examining. The host adds that valuations are already very low and thinks selling pressure from cashing out may ease after the offshore tax-related deadline on October 22. Rick also points out that overseas flows remain constrained by dollar rates.

8. In his framework, AI borrowing leads to Treasury yields above 6%

• Rick estimates AI-linked companies will issue about $300B in debt in 2026, adding financing demand equal to roughly 20% of U.S. government borrowing. If it keeps growing, the figure could reach 30% in 2027. He stresses that this is new incremental demand and admits he cannot recall the exact figures and has not “worked them out in detail.”
• The chain of cause and effect: demand for funding rises, supply does not rise alongside it, and rates get pushed up. He expects the 10-year Treasury yield to “reach above 6 sooner or later.” Six percent may not hold as a ceiling; 7% or 8% would not surprise him. “Higher for longer” should become “much higher.”
• He compares the situation with China during reform and opening up, when real estate and major infrastructure projects drove lending rates into the high 8% range and informal borrowing rates into the 20s. Measured against GDP, he thinks AIDC construction could require more incremental funding than China’s infrastructure buildout at the time, so high rates may not fade quickly.
• If 10-year yields stay above 6% for a long time, he fears dollar funds will flow back from Asia and other markets, setting off a crisis like the 1997 Asian financial crisis, the eurozone sovereign debt crisis or a Latin American crisis. “I don’t know where it starts, but it will happen.” Real estate will also face high rates and the elevated unemployment he expects.

9. He has trimmed Nvidia and is getting more selective within AI

• Rick says his view has become more extreme: he is moving from a “K-shaped” market to an “I-shaped” one, choosing only the strongest assets even within AI. Broadcom and Qualcomm are examples of fabless businesses he is wary of, though he still says he owns Qualcomm.
• He says he sold most of his Nvidia shares about two months ago. The main reason he has kept the rest is that he has owned the shares for less than a year and would face a higher tax bill if he sold. His view is clear: “CUDA’s moat no longer exists”—or at least it is nowhere near as wide as it used to be.
• The logic is that model companies are all working on chips, and the next step is to iterate on chip design itself. Nvidia’s plug-and-play, large-scale systems have value, but Google, Amazon and Anthropic could build similar systems. If design cycles keep shrinking, chip designers could face the same pressure on value as software companies.

10. Manufacturing, equipment and optics have moats built on time

• Compared with design, Rick prefers TSMC, Intel, ASML and core equipment makers. For test equipment and FORMFactor probes, once a supplier is qualified for a production generation, replacing it often means starting qualification over. A live production line cannot absorb the cost or accident risk of switching casually. The moat comes from “time that can’t be shortened.”
• In optics, he bought Coherent after it fell, along with upstream equipment supplier Veeco, which makes MOCVD and indium phosphide growth equipment. He is still on the fence about Lumentum and has not bought it.
• By his comparison, Aixtron is the largest player in this equipment market and Veeco is number two. Customers do not switch equipment once it is installed, especially when there are few suppliers and capacity is still tight. These are long-term positions; he is not expecting an immediate catalyst to deliver a huge move.

11. Memory remains his largest position; he rejects the call that the cycle has peaked

• Rick says flatly, “I don’t know when it’ll come back,” but sees a large margin of safety in memory. He has not changed his share positions, though memory options have lost a lot. He has not started buying memory options back; for now, he is buying the market’s least-loved sector, NeoCloud.
• He cites reports that supply is already sold through 2028 and 2029, using SK hynix as an example. For retail demand, he shares a personal anecdote: he ordered, then canceled, two 64G memory sticks, 128G in total, and they rose about 50% not long after. He offers this as a concrete rebuttal to the idea that demand has peaked.
• The core mechanism is that model improvements increase the intelligence and potential revenue produced by each unit of compute. The same card could run a stronger model later, making the hardware more valuable. Asked where memory, GPUs, and H100, H200 and GB200 rental rates will eventually settle, he says, “I don’t know.”
• He thinks the “cycle” thesis is wrong: demand growth has not stopped, and the growth rate “very likely hasn’t slowed down.” Even if acceleration along part of an S-curve declines, a new model could accelerate demand again. If the market is right and he is wrong, he accepts the loss; his bet is that current prices undervalue this capability curve.

12. He already owns test and packaging names; AEHR is only a conditional small bet

• Across materials, testing and packaging, Rick mentions Applied Materials, KLA, AMKR, FORMFactor and Teradyne. He owns fewer Taiwanese stocks and names only UMC specifically. He says he has positions in these areas, but not large ones in every name.
• AEHR is a small position he bought after another KOL recommended it. If its wafer-level burn-in process is accepted by SK hynix, Micron and others and becomes standard for HBM, the stock could rise substantially. He admits he knows little about the technology and calls it a high-risk, high-return bet.

13. The NeoCloud debate centers on expansion financing and valuing power contracts

• The host relays a report that Elon Musk may be preparing to buy a cash-strapped NeoCloud, with a claim involving “more than 200” companies. Rick says NeoCloud is the sector where he has lost the most; Bio used to be his second-largest source of losses, but has mostly recovered.
• Using a company he calls “Coinbase” as an example, he separates existing debt from future expansion. Existing financing has a locked-in rate and revenue is also locked in; higher electricity costs can be passed through. More expensive new borrowing would impede expansion, but he does not think existing debt alone is enough to cause bankruptcy: “It’s a landlord’s business.”
• Power contracts are a separate part of the valuation. He cites about 1.5GW already in operation and another 1.5–2GW signed and queued. If the company sold about 1.5GW of those contracts, he thinks it could get back several billion dollars. After including the value of those unbuilt contracts, an AI tool put its price-to-book ratio at about 1.6–1.8x. That is his adjusted estimate.
• He thinks a bond yield of about 15% reflects excessive bankruptcy concern, but he does not rule out failure: management could make mistakes, and power contracts could lose value, taking those multiples back to 2.6, 2.8 or even 3. He is betting on another path: contracts keep rising in price, adjusted price-to-book falls below 1, and the company then finds financing easier to secure.

14. He sold all his Bitcoin because of uncertainty around math and quantum risk

• Rick confirms he has sold all the Bitcoin he held for years and plans to trim some Ethereum as well. He will still hold some crypto to hedge against fiat risk. This is an actual position change, not just an expression of concern.
• He still agrees with Arthur Hayes’s long-term optimism on crypto and thinks an agent economy can only be built on crypto. But “which assets benefit” remains uncertain. Relative to Bitcoin, he prefers Ethereum, so for him the long-term case for the technology and the choice of which asset to hold are separate judgments.
• The first layer of risk comes from AI advancing mathematics. He recounts Vitalik’s concern that the assumptions underpinning cryptographic security could be undermined and discusses the Exponential Time Hypothesis—not Ethereum’s ETH. On P versus NP, he still thinks P will probably be proved not equal to NP eventually, but worries that the opposite result could threaten cryptography’s foundations.
• The second layer is AI accelerating quantum computing. He does not know whether breaking elliptic-curve signatures really requires 1,000–1,500 logical qubits, or whether those capabilities are 5–7 years away; he thinks it could be just 2–3 years. These are timeline risks he no longer wants to carry in a concentrated position.

15. Even if Q-day never arrives, fear could move prices first

• Rick stresses that even if the technical risk never materializes, concern alone could drive a steep decline. Bitcoin’s liquidity is small relative to its market cap, so a modest wave of selling could have a large price impact. He would rather sidestep the “squat before the jump”; he is more worried that the asset he holds during the dip may not be the one that jumps afterward.
• The host adds a conditional scenario: if public keys could be used to derive private keys, control over funds would be lost. Responses could include no longer recognizing assets from before a certain block or upgrading every address to be quantum-resistant. As he recounts it, those choices could conflict with Bitcoin’s decentralized ethos.
• The host recalls Sun Ge giving a window of about one to one and a half years and mentions a TRON upgrade, but also wonders whether he is trying to “scare everyone, then buy the dip,” adding, “I don’t know.” Rick is also unclear on TRON’s specific plans. His conclusion is still to avoid this highly uncertain period for now.

16. He is staying invested while waiting for several technology curves to compound

• Asked about the IPOs in the week of November 16, year-end and Q1 2027, Rick answers, “I don’t know,” and admits his earlier call for a V- or U-shaped reversal was wrong. He is willing to endure a 20%–30% drawdown, even followed by another 30%. In his recollection, missing the single strongest month of NTC’s run meant missing 70% of its gains.
• He compares the current pullback with the end of last year, when Micron fell from about 240 to 190 and the market questioned whether prompts and workflows could justify capex. Coding later provided a use case. The question now is “What is there besides coding?” He sees AI for science shaping up as the next use case, though the pullback could last longer. He does not want to bet on being able to time it.
• His longer-term view is that AI will compress the usual technology cycle and bring successive waves of progress in life sciences, materials, energy, fusion and quantum computing. He recounts Dario’s expectation that AI will solve all cancers in the next 7 years, and infers that the rally could be more violent while the cost of waiting could be lower. This is still the path he is willing to bet on.
• His current holdings, from largest to smallest, are memory, chips, Bio and equipment. He names Micron, SK hynix, Samsung, NTC, ASML and TSMC. “The way I held Bitcoin back then is how I’m holding the AI sector going forward.” His rule is “time in the market, not timing the market,” as long as the core thesis has not reversed.