GoodAlexander - The AI Trade Is Topping, Google’s Search Problem and the Long Coins Trade

Summary

• GoodAlexander says the AI trade is topping as inference growth slows and the ROI story weakens. Anthropic’s inference growth fell from 140% to 45% quarter on quarter; investors expect Fable 5.5 to reaccelerate Q4. He estimates adjusted inference gross margins at 50–60%, but a DeepSeek pro model matching Opus could push them to 10–15%: “That’s a disaster” at these valuations.
• Google’s search problem is losing monopoly pricing while AI investment consumes cash. Thread Guy cites a 13.5% selloff and search falling below 50% of revenue. GoodAlexander sees competing AI interfaces weakening monetization despite strong consumer spending; his challenge to YouTube’s growth story is pointed: “Why are you growing slower than fucking Cheesecake Factory?”
• CoreWeave’s 15% parent debt does not represent the financing cost of its entire GPU business. GoodAlexander bought debt because asset- and tenant-backed financing can clear around 6–9.5%, making the blended hurdle more manageable. The vulnerability is rental prices: if everyone expands into slowing inference demand, “we’re just gonna saturate the market with tokens.”
• Failed AI investment worries him more than a wave of AI layoffs. Accenture still has around 800,000 employees, and AI deployment itself requires hiring; his recession risk runs through disappointing returns, data-center retrenchment and weaker consumption. That caution extends to semiconductors: “The downside of tech up here is very, very high” if investors start valuing Nvidia as cyclical.
• Hyperliquid’s rerating rests on traditional finance moving on chain. TradeXYZ represents roughly 50% of its volume, varying daily, while crypto revenue growth itself remains unexceptional. The appeal is “early innings of blockchains becoming a TradFi settlement layer”; he prefers fee-generating HYPE and BNB to ETH and SOL, although HYPE’s rally has reduced its appeal.
• His long-coins thesis combines a liquidity rotation with fiscal and privacy demand. “Coins have bottomed out,” he says, while AI is losing its pull on capital. BTC benefits in his framing from roughly 6.5% annual US deficit spending and potential policy rescues; Monero benefits from triple-digit DeFi hack growth and easier access through Hyperliquid.
• Flat stablecoin supply limits his enthusiasm for ETH and SOL, while a blue wave threatens regulatory progress. Circle’s velocity explanation does not satisfy him: “We need to see the number go up.” Under a blue wave, he expects Elizabeth Warren to become banking chair and scrutinize Trump-family crypto ties—bad for integration, but potentially supportive of capital-flight demand for BTC and privacy coins.

Deep dive

1. Hyperliquid is being valued as settlement infrastructure

• GoodAlexander reads Stan Druckenmiller’s TradeXYZ endorsement as support for a structural rerating: roughly 50% of Hyperliquid volume comes from TradeXYZ, varying daily, and traditional assets drive growth. Crypto revenue itself is not growing especially fast year on year. The appeal is being “in the early innings of blockchains becoming a TradFi settlement layer,” with Canton’s fixed-income story another expression.

• The Robinhood comparison: overall revenue grows around 12%, while prediction markets and index zero-DTE options grow much faster. Customers favor “2% odds for 50x payouts.” But monthly users are decelerating, and the mobile brokerage revolution has happened; Hyperliquid offers investors a new infrastructure story.

2. Inference deceleration puts frontier valuations under pressure

• The Anthropic S-1 mattered less to GoodAlexander than investor-day disclosures: inference growth fell from 140% to 45% quarter on quarter, with no Q4 guide. Opus 5.5 is “a tasteful and very good model”; investors expect Fable 5.5 to reaccelerate Q4. If inference decelerates, even the valuation’s retreat from $2 trillion to $1.5 trillion may be insufficient.

• Vibe coding’s slowdown is an ROI problem in his telling: people thought they would make an app that made money, but didn’t. Traditional software engineers are now the people using AI tools. NPM download growth fell from triple to double digits. ServiceNow billings weakened as it dropped FDEs; Fortinet guided Q4 down and struggled to pass on memory costs, contradicting the cybersecurity-emergency pitch.

• Previously excited by inference growth, he became more cautious about OpenAI’s cybersecurity pitch. If powerful models force wholesale software rewrites, expenses rise without corresponding revenue. His Zoom example: customers still pay the same after a complete rebuild, so “Zoom’s margins are gonna go off a cliff.”

• He rejects the claim that inference is given away free: AWS showed an 80% GAAP gross margin on inference, and after adjusting for converts, Thread Guy estimates margins might still be 50–60%. He says Amazon’s Anthropic shares are behind subsidized compute deals. But if a new pro model matches Opus at a much lower cost, margins might fall to 10–15%—“That’s a disaster” at these valuations.

3. Google’s search weakness exposes the gap between AI and earnings

• Thread Guy acknowledges bad earnings but finds Google’s selloff jarring: 13.5% on a roughly $4 trillion company, with search below 50% of revenue for the first time. GoodAlexander contrasts single-digit Q4 AdWords growth with Cheesecake Factory’s 15% comp: strong consumer demand makes advertising’s deceleration harder to explain away.

• His answer: AI search monetizes worse because a monopoly has become a four- or five-player interface market, including ChatGPT with around a billion users. That reduces pricing power. Google, he says, would rather discuss AI progress than search numbers while going free cash flow negative for the first time.

• The market allows Elon to accelerate SpaceX spending because he has a track record of getting things done, relationships in Texas to build data centers, and has rented capacity to Anthropic. Physical AI supplies his distinctive pitch: a dolphin made 10 times smarter still “doesn’t have opposable thumbs.” GoodAlexander says physical AI training data is difficult or perhaps impossible to distill, describing individual data sales worth $5–10 million.

• China is his counterexample to AI automatically helping earnings: a property wealth shock has crushed consumption despite strong models. Mooncake sales are down around 65% year on year; Hang Seng Tech trades around 10 times EBITDA versus 23 times for the Nasdaq. “You might have great AI models,” but if the consumer rolls over, “it’s all downside.”

4. Neocloud financing works only while rental economics hold

• GoodAlexander says CoreWeave’s operators were Ethereum miners and had personally sold a couple billion dollars of stock over the last year before the company accelerated borrowing and investment. A roughly $100 billion, five-year backlog does not prevent near-term cash burn, even with three-year paybacks. Investors face opposing risks: acceleration brings competing chips and depreciation; deceleration undermines inference demand. Neither camp wanted the financing.

• GoodAlexander bought the debt because the 15% headline ignores its structure: roughly 70% of a deal can be financed against GPUs and tenant commitments, such as Microsoft’s, around 9.5% or less, potentially 6% for a high-quality tenant. The last 20% uses parent debt yielding 15%. Vera Rubin looks capable of clearing the ROI hurdle, assuming market prices and rental rates hold.

• His concern is that assumption: he says a GPU can generate $30 an hour running DeepSeek while rental costs are $10 an hour. If everyone chases that spread while inference decelerates, “we’re just gonna saturate the market with tokens.” Today’s attractive prices need not persist.

• Oracle’s 14% rebound rewards investment cuts and a cap of $10 billion each on additional stock and debt. Management still claims two-year paybacks, free cash flow breakeven in 2028 and $30 billion of free cash flow in 2029. Credit-default-swap spreads are falling; CoreWeave equity still risks dilution if it must finance deals in a weak market.

5. The employment risk runs through disappointing investment returns

• Thread Guy raises Microsoft’s 60,000 buyout offers after Block’s layoff announcement lifted its stock 15%. GoodAlexander pushes back: Accenture still has around 800,000 employees, and AI transformations require forward-deployed engineers. Commodity software workers are exposed, but “the idea that there’s mass unemployment is clearly false.”

• Bank of America illustrates the scale problem: he cites $300 million of Mythos savings, rising to $700 million next year, against a $2 billion money-laundering charge and 6,000 compliance hires. His bigger employment worry is failed AI CapEx, data-center retrenchment and negative GDP, turning Cheesecake Factory’s positive 15% comp into negative 15%.

• For SOXL and semiconductors, he flags a January Huawei–DeepSeek training-chip launch. Vera Rubin upgrades look exciting assuming inference demand holds. But Korean equities trade at six times earnings; repricing expensive Nvidia as cyclical could be painful. Asked whether buying semiconductors here offers good risk/reward: “No, I think it’s terrible.”

6. Long coins means selecting fees, fiscal hedges and privacy

• His preference is fee-generating HYPE and BNB over ETH and SOL, although HYPE is less compelling after its run; BNB still looks reasonable. “You generally just want to own the cheapest fee-generating assets you can on-chain.”

• Monero’s specific demand driver is noncyclical, triple-digit year-on-year DeFi hack growth. The Wagyu bridge lets Hyperliquid users reach spot XMR without KYC, opening incremental demand. GoodAlexander keeps the operating uncertainty explicit: “We’ll see if it stays open.”

• BTC benefits in his framing from roughly 6.5% annual US deficit spending and possible fiscal and monetary action if growth weakens. AI was sucking liquidity out of crypto; now, he says, “coins have bottomed out.” If crypto hacks decelerate while TradFi hacks accelerate, crypto’s forced security adaptation could become an advantage.

• His counterweight is the stablecoin supply chart: Tether and Circle aren’t getting incremental volumes. Circle’s velocity explanation does not satisfy him: “We need to see the number go up.” He says materially increasing stablecoin supply would make him more excited about ETH and SOL.

7. Political backlash could split crypto’s winners

• The next two years could entrench crypto in TradFi, making it harder to unplug. Under a blue wave, however, he says Elizabeth Warren would become banking chair and the focus would turn to World Liberty Financial and Trump-family ties. He worries about integration while expecting capital-flight demand for Bitcoin, Zcash and Monero.

• His final political call draws on what he describes as Palmer Luckey’s genetically engineered war dogs: religious voters will turn against AI over “playing God.” With the pope already against it in his telling, he thinks that backlash probably puts a damper on the political right in the US and globally.