China just K*LLED Claude AI.. - GoodAlexander & David Choi
Summary
GoodAlexander argues Chinese open-source models have broken the economics underpinning Claude and other closed-model leaders. After initially dismissing them as benchmark-gamed Opus distills, he tested models delivering Opus- or Sonnet-like performance on a roughly $10,000 GPU, even 3090s. Against protected Opus API usage at about $80 per terminal-hour versus roughly $2 for Qwen 3.6-class compute, the “40 to 80x delta” makes local inference and crypto-financed compute newly credible.
David Choi sees an enormous financing gap beneath the AI-capex boom: Nvidia has an estimated $500 billion of sales over the next five quarters, “99% of that is financed,” yet loans below roughly $200 million barely exist. He says 20-30% of those sales come from smaller emerging clients that cannot efficiently access securitized debt. USD.AI’s next-12-month target is $1 billion of GPU loans—enough, he hopes, to establish a global reference rate for the cost of growing AI.
The core USD.AI pitch is that GPUs should be financed like houses, cars, and aircraft, but their product cycles move too quickly for traditional securitization. Packaging 10,000 loans, tranching them, conducting a roadshow, and selling the instrument can take two years—long enough to move from Hopper through B200, Vera Rubin, and toward an unannounced Feynman generation. DeFi instead adds each originated loan immediately to the sUSDai NAV, creating what David calls “a debt perp.”
USD.AI’s proposed flywheel uses stablecoin economics to lower AI’s financing cost, not merely to put existing debt on-chain. If AI companies settle in USD.AI, the protocol can use the 3-4% earned on each minted dollar to subsidize borrowers while retaining origination fees and net interest margin. CHIP is framed as the protocol’s GP-like governance layer, directing fees, helping set rates across deal types, and potentially prioritizing redemptions—not as a fully disclosed utility package yet.
Compute scarcity is presented as both the reason Claude feels worse and the foundation of a new sovereignty contest. David says providers are rate-limiting users because demand is outgrowing installations; OpenAI’s aggressive compute investment was therefore directionally right. His “AI dollar” thesis is that countries will become net importers or exporters of generated tokens, and a nation without sufficient AI capacity will fall behind much as an oil-poor country did during the petrodollar era.
GoodAlexander’s crypto bull case spans three different clocks: Saylor-driven inflows in one year, capital flight in three, and crypto-native financial rails in five. STRC can bring retirees’ capital into Bitcoin, while failed attempts to suppress privacy coins demonstrate that privacy technology may survive a hostile policy regime. Longer term, taxation, money printing, CBDCs, and capital controls could make portable crypto assets attractive to wealthy people who cannot “get on a plane with a bag of gold.”
His equity view is almost the mirror image: stay bullish into the AI-IPO spectacle, then prepare for a historic reversal. He expects roughly $4 trillion of SpaceX, OpenAI, and Anthropic equity to come onto the market, with a $1.5-1.7 trillion SpaceX IPO as “the main show.” If banks support favorable AI comps and passive indexes rapidly absorb SpaceX, he sees that event as the “Solana $300 moment”—the beginning of a generational top rather than a permanent new plateau.
Deep dive
1. Compute—not model intelligence—is becoming AI’s binding constraint
David’s opening claim is that Claude gets “knee-capped every time you try to use it” because providers lack compute, not because model research has stalled. In the long run, he argues, the decisive question is simply: “How many Blackwells do you have access to?”
The speakers use a physical comparison: build 30 Las Vegas Spheres and, by their estimate, “that’s one data center.” They put one gigawatt at roughly $50 billion and describe planned sites reaching two or three gigawatts across dozens of global projects.
David says AI capex is now so large that, absent it, US GDP growth would be negative 3-4%. He further claims one year of AI infrastructure spending can exceed a decade of spending on oil, highways, and other conventional infrastructure.
USD.AI joins what he calls two megatrends: stablecoins, among the world’s fastest buyers of Treasury bills, and trillions of dollars of AI capex. The thesis is that the “fastest horse” in capital formation should finance the economy’s fastest-growing capital requirement.
2. Chinese open models have punctured the closed-model consensus
GoodAlexander says the prior consensus treated Chinese releases as stolen Opus distills that Anthropic would eventually shut down. He shared the skepticism—Chinese labs can “over-promote” releases and game benchmarks—but changed his mind after personally testing the latest models.
The break is hardware accessibility. Workloads that recently required elite closed APIs can now, he says, achieve Opus- or Sonnet-like results on a $10,000 graphics card, sometimes on 3090s: “Prior to two weeks ago, doing that on commodity consumer hardware was not really viable.”
His protected Opus setup costs roughly $80 per terminal-hour, or $160 while running two terminals, versus around $2 per hour for Qwen 3.6-class alternatives. That “40 to 80x delta” led him from “OpenAI has just won” to believing open source remains seriously underpriced.
In their discussion of newer models, GoodAlexander points to GLM 5.1 and Knowledge Atlas Technology (ticker 2513). David says GLM is, in his view, a better chat model than Opus; both criticize Claude/Opus for excessive verbosity and Anthropic’s overcorrection from sycophancy.
GoodAlexander connects the shift directly to crypto AI: closed models such as OpenAI and Anthropic are gated and unlikely to route meaningful compute spending through crypto, while open-source models make distributed, crypto-financed inference plausible. He expects explosive growth in Hermes Agent and other open-source harnesses.
3. GPU credit resembles housing finance before modern mortgages
GoodAlexander traces David’s edge to an earlier lending system for NFTs—the “worst credit in the world,” borrowers financing internet profile pictures. GPUs are warehouse-finance assets with an asset-backed-security element that he considers superior to private credit’s current treatment.
David’s mortgage analogy starts with liquidity. Banks can offer 90% house financing because thousands of otherwise illiquid mortgages are packaged into tradable securities; no equivalent secondary market exists for a single GPU, and “if you can’t get a loan for a single GPU, you can’t get a loan against a large cluster.”
Traditional issuance is too slow for the hardware cycle. By the time loans are accumulated, tranched, marketed, and sold over roughly two years, the collateral may have progressed from Hopper to B200, Vera Rubin, and toward Feynman—“the ones they haven’t even announced yet.”
Below roughly $200 million, David says institutional GPU financing effectively disappears. One emerging data-center operator received a five-page Guggenheim or Jefferies explanation for why its $200 million purchase could not be financed; private credit is retreating, while Oracle is portrayed as already “max levered out.”
4. Expensive equity fills the gap that asset-backed debt should occupy
Without dedicated GPU credit, borrowers resort to convertibles that can become equity. David’s example: early CoreWeave or Crusoe holders may suffer roughly 90% dilution because issuance becomes the only available way to fund equipment.
GoodAlexander calls this the “Bitcoin pizza dynamic.” If a company sells equity at a $1 billion valuation to buy $100 million of GPUs and is later acquired for $60 billion, it effectively surrendered about $6 billion of expected value for that hardware.
His Finance 101 conclusion: a founder bullish on the company should preserve equity and finance the asset directly. GPUs also carry large tax value because they are depreciable capex; David cites Meta’s roughly $20 billion of revenue alongside $20 billion of capex as an almost deliberate pairing.
David points to bonus depreciation and other tax incentives as an important part of the yield economics. The speakers’ broader argument is that cheaper GPU financing avoids dilution while preserving the depreciation benefits of the underlying buildout.
5. DeFi turns slow securitization into a continuously priced “debt perp”
David’s answer to “why crypto?” is speed of capital formation: DeFi can “spawn markets out of nothing.” A roughly $1.5 billion CoreWeave bond had traded only about $50 million since August, while USD.AI’s much smaller system allegedly produced around $15 billion of secondary volume in six to eight months.
The multiplier comes from composability. Once USD.AI and sUSDai exist, users can trade them, strip yield through Pendle, loop positions through Fluid, and create implied notional exposure far beyond the original loans.
Each new loan can enter the debt instrument or sUSDai’s NAV immediately rather than waiting for thousands of assets and a roadshow. David contrasts Hyperliquid’s futures perpetuals with USD.AI’s ambition: “We’re really like a debt perp,” continuously pricing and compounding newly originated credit.
The structure is intended to broaden access beyond white-collar bond desks. David’s claim is not merely that blockchains distribute old securities more widely, but that immediate pricing and settlement make a previously unavailable small-ticket GPU market possible.
6. USD.AI wants to discover—and eventually subsidize—the interest rate of AI
Chapter one is explicit: originate about $1 billion of loans over 12 months, becoming the reference rate for GPU credit. To David, that rate is “how fast you can grow AI”—the cost of capital and the amount people will spend for a return.
He expects ownership to spread down the stack. Rental providers are buying chips; token factories that amplify output per GPU want ownership; LLM companies already buy; applications will follow because indefinitely renting their largest cost is too expensive.
David puts compute at roughly 80% of OpenAI’s cost. He says $500 billion of Nvidia equipment becomes $600-700 billion after interest, making Wall Street effectively AI’s second-largest business behind Nvidia and ahead of OpenAI.
If customers settle in USD.AI, its 3-4% income from minted dollars can subsidize debt rates. Revenue still comes from origination fees and net interest margin. GoodAlexander frames CHIP as governing fee routing, rate decisions by deal size and off-take quality, and possible redemption priority; fuller utility details remain forthcoming.
7. Real customers are being pulled on-chain by financing, not ideology
David reports about 70 incoming customers, including large neoclouds, with demand increasing daily. The immediate constraint is enough protocol liquidity to serve them, hence the push to deepen sUSDe and USDe markets after CHIP’s launch.
Many customers have never used a wallet. The team introduces them to Coinbase Prime and, where European access failed, even sent free Ledgers; initial confusion gives way to amazement that they can borrow against GPUs, buy more GPUs, and repeat through a smart contract.
The founding insight came while David mined Bittensor. He says paying about $3 an hour for a rented chip amounted to more than $25,000 annually, while the GPU itself cost roughly $25,000: “Imagine if you bought a house for a million dollars, but your rent is 100k a month.”
The host disclosed that he previously held and traded CHIP, made money, and held none during the discussion; he remained cautious on crypto alts. David’s own emphasis was operational: “I’m just doing more loans,” with most fees expected from AI companies rather than crypto users.
8. AI capacity is becoming a sovereign resource and potential monetary anchor
David expects jurisdictions to become net importers or exporters of AI tokens, meaning generated inference rather than crypto tokens. A country’s token-production capacity will increasingly determine citizen productivity, economic output, and strategic autonomy.
His petrodollar analogy begins in the 1970s: oil powered cars, heating, factories, and national growth, so controlling its settlement conferred monetary and geopolitical leverage. AI now plays the analogous role because a citizen equipped with it may have “10x” the capacity of one without it.
David states the conclusion categorically: “If you don’t have AI, you’re not a country.” The extraordinary data-center buildout is therefore not just corporate capex but an arms race to keep the underlying resource continuously available at massive and smaller scales.
9. The crypto bull case rests on inflows, censorship resistance, and capital flight
GoodAlexander moved from mocking Michael Saylor’s promotion to embracing Taiki’s STRC thesis. He believes Saylor has “cooked this inflow machine”: an approximately 11%-yielding instrument aimed at retirees that turns older savers’ capital into mechanical Bitcoin purchases and can expand MicroStrategy’s mNAV flywheel.
He supports the demographic point with Asian retail behavior. Crypto Twitter is about 93% male, he says, while Japanese and Korean participation is nearer 50/50; older Korean women are meaningful speculators because years of weak local equities and popular forex trading produced a different investing culture.
Despite exchanges targeting Monero and pressure around privacy projects, GoodAlexander says Monero and Zcash have continued to work; Zcash’s presence on Robinhood reinforced his view that “you try to kill these privacy coins and you can’t.”
His horizon map is concise: one year of Saylor-supported inflows, three years of capital flight, and five years for the future to be built on crypto rails. Money printing, wealth taxes, CBDCs, and capital controls push wealth outward; unlike gold or oil, crypto can cross a border without a physical bag.
His long-run macro frame is darker: he believes US governance and property-rights enforcement have deteriorated, implying an eventual 11-13x earnings terminal multiple rather than roughly 29x. He says living in Puerto Rico and keeping wealth outside the system are ways to remain robust through that transition.
10. The AI-IPO supercycle may end by turning passive equity into high-FDV crypto
The host compared current markets to Solana’s “daily-runner” era: one speculative ball ricochets among Car, Hims, Texas Instruments, and other names, while the S&P’s relentless passive inflows distinguish it from crypto. GoodAlexander’s answer is that the frenzy can persist until its largest liquidity event arrives.
David estimates that SpaceX, potentially at $1.5-1.7 trillion, is “the main show,” with OpenAI and Anthropic following. He is more bullish on Sam Altman’s ability to engineer narrative and stock price than on Dario Amodei, and expects the sequence to be SpaceX first, OpenAI second, and Anthropic last.
David says banks and companies are incentivized to upgrade AI stocks and engineer impressive enterprise deals because weak comparables would threaten enormous IPO fees. He gives possible examples such as JPMorgan doing an enterprise deal with Cursor or Cognition selling a “kajillion” seats.
The danger arrives if SpaceX is rapidly added to the S&P 500 or Nasdaq, forcing retirement portfolios to buy it without company-level diligence. David calls that “imagine if EOS got added to the S&P 500”: a high-FDV crypto launch embedded in passive finance, followed potentially by a “generational top” and “the mother of all shorts.”