Featuring
Alan Keating joins for the closing poker clinic, but the sharpest action comes before the cards: Tether's trust deficit, AI-chip depreciation, and Google's Search defense. Jason has the best episode by landing the transparency burden on Tether and calling Google's near-term growth, while Friedberg turns Accounting Corner into an actual win—a feat with longer odds than Keating's four-deuce.
Spice rack
Has Tether cleaned up enough to deserve trust, or does its compliance history still justify skepticism?
Original point: Tether has improved, but its history of weak reserve transparency, jurisdictional restrictions, and illicit use makes continued scrutiny necessary.
What everyone argued
Chamath Palihapitiya
Chamath says Paolo Ardoino presented as a credible operator running an extraordinary business: hundreds of millions of people get dollar access, Tether invests the backing in Treasuries, and the spread produces enormous profit. He pushes Jason not to turn criminal use of USDT into an accusation against Tether without evidence of issuer complicity.
Jason Calacanis
Jason credits Tether's recent cleanup but argues that the company earned skepticism through years without a full audit, reserve controversy, and restrictions in important markets. He says the new regulatory framework can force stablecoin businesses into clearer, more accountable structures.
Winner circle
Jason wins narrowly on the burden of proof. Tether's business strength and reserve attestations are real, and Chamath is right that criminal use of USDT does not prove issuer complicity. But Tether's own 2026 announcement confirms that a full financial-statement audit remained unfinished, so its history still justified scrutiny rather than a trust victory lap.
Commentary
Chamath Palihapitiya
Assumptions and fact checks
Rapid adoption and a highly profitable Treasury-backed model are strong evidence that Tether is now institutionally trustworthy.
Why it mattersScale, liquidity, and profitability reduce some business-model risk, but they do not substitute for controls testing, consolidated financial statements, or a completed independent audit.
Evidence that criminals used USDT does not by itself show that Tether participated in their crimes.
Why it mattersTreasury has documented illicit networks using USDT, but use of a payment asset is different from proof that its issuer knowingly joined the scheme.
Tether had surpassed 500 million users and had roughly $135 billion of direct and indirect U.S. Treasury exposure by Q3 2025.
CheckTether's Q3 2025 release, based on a BDO attestation for reserves and company analysis for users, reported more than 500 million users and approximately $135 billion of Treasury exposure. The reserve figures received attestation assurance; the user count is company-reported rather than a financial-audit opinion.
Tether had generated more than $10 billion in profit during the first three quarters of 2025.
CheckTether's Q3 release reported year-to-date net profit above $10 billion alongside the BDO reserve attestation. It was not a full audit of Tether's consolidated financial statements, so the assurance scope matters.
Jason Calacanis
Jason wins the transparency burden but loses precision on criminal use. The sharper case is already strong: an issuer with official findings for misleading reserve claims should complete a full audit before asking critics to stand down.
Assumptions and fact checks
Tether's earlier reserve misstatements and audit gap justify a higher continuing burden of proof.
Why it mattersOfficial findings and the absence of a completed full audit make enhanced scrutiny reasonable even after later attestations and compliance improvements.
Regulatory access in the United States is a reliable shorthand for the safety of the global USD₮ business.
Why it mattersU.S. compliance is useful evidence, but it does not cover every reserve, affiliate, customer, or jurisdictional risk in Tether's global structure.
New York barred Tether and Bitfinex from further trading activity with New Yorkers and imposed an $18.5 million penalty.
CheckThe New York attorney general's 2021 settlement required the companies to cease trading activity with New Yorkers, pay $18.5 million, and make additional reserve disclosures after findings of misleading reserve statements and concealed losses.
Tether did not have a full independent financial-statement audit at the time of the episode.
CheckTether published periodic BDO attestations, which are narrower than a full financial-statement audit. In March 2026, Tether announced an engagement for what it explicitly called its first full independent financial-statement audit.
Tether was involved in human trafficking.
CheckOfficial sources document illicit actors using USDT for sanctions evasion, laundering, and scams, but that does not establish that Tether itself participated in human trafficking. The statement conflates use of the token with issuer complicity.
Do six-year AI-chip depreciation schedules reflect real useful life or flatter big-tech earnings?
Original point: Michael Burry's accounting accusation is wrong because companies may depreciate chips across the years they remain in productive use, while investors can see the cash spending and choose their own valuation measure.
What everyone argued
Chamath Palihapitiya
Chamath says chip value must be tied to the revenue produced by output tokens, not treated like undifferentiated energy. Providers meter usage, gate uneconomic demand, and rebuild serving infrastructure, so older chips can keep creating valuable output even as new hardware arrives.
Jason Calacanis
Jason argues that an H100 may deliver most of its consumer and economic value in its first three years, then spend years on smaller background jobs. Straight-line accounting can therefore be legal yet too coarse for hardware whose productivity and strategic importance fall sharply by vintage.
David Friedberg
Friedberg says GAAP estimates useful life from productive use rather than the arrival of a better model. If old chips still generate revenue, their life can extend; if new equipment displaces them or performance forces retirement, companies must revise estimates or recognize impairment. He adds that capital spending is visible in the balance sheet and cash-flow statement, so investors are free to value free cash flow instead of accepting GAAP earnings at face value.
Winner circle
Friedberg wins the accounting round, with Jason preserving the investor caveat. Continued productive use, periodic reassessment, and disclosed cash spending make a six-year life defensible rather than inherently fraudulent. But six years is an estimate, not a law of nature: investors should test it against cohort utilization and retirements, and companies should shorten it when the evidence changes.
Commentary
Chamath Palihapitiya
Chamath improves the debate when he asks what each output token earns and degrades it when he calls Burry bad at his job. The accounting estimate should be beaten with cohort economics, not contempt.
Assumptions and fact checks
AI providers can measure output-token economics well enough to support long server useful-life estimates.
Why it mattersProviders have detailed internal cost and workload data, but shared infrastructure, bundled subscriptions, free usage, model training, and rapidly changing serving stacks complicate asset-level attribution.
A chip that still produces some revenue remains economically useful on roughly the same basis as in earlier years.
Why it mattersUseful life does not require constant productivity, but residual revenue alone cannot establish the proper life or straight-line pattern; margins, maintenance, displacement, and workload mix matter.
Jason Calacanis
Jason has the best losing objection: a server can remain busy after its premium workloads move elsewhere. He needed fleet-vintage evidence instead of a round 90% estimate to turn that insight into an accounting verdict.
Assumptions and fact checks
AI accelerators deliver roughly 90% of their lifetime value in the first three years.
Why it mattersThe episode supplies no cohort utilization, revenue, margin, resale, or retirement data for that number. Different chips and workloads can age very differently.
Straight-line depreciation is inherently misleading for rapidly improving AI hardware.
Why it mattersIt can be coarse when benefits are front-loaded, but the more important question is whether the total useful-life estimate and impairment reviews reflect actual retirement and workload evidence.
NVIDIA's fiscal Q3 2026 revenue rose 62% year over year and 22% quarter over quarter to about $57 billion, while GAAP net income reached about $31.9 billion.
CheckNVIDIA reported $57.006 billion of revenue, up 62% year over year and 22% sequentially, and $31.910 billion of GAAP net income. Its following-quarter revenue outlook was $65 billion.
Large technology companies can revise server useful lives downward when operating evidence changes.
CheckAmazon's 2025 filing says it shortened the estimated life of a subset of servers and networking equipment from six years to five after extending servers from five to six in 2024.
David Friedberg
Friedberg wins the accounting question because he explains where judgment lives and where cash can be checked. The honest finish is narrower than his victory lap: Burry has not proved fraud, but investors should still test six-year lives against retirement and workload data.
Assumptions and fact checks
Continued revenue generation is enough to validate the selected useful life.
Why it mattersContinued productive use supports a nonzero remaining life, but the estimate also depends on expected retirement, workload displacement, maintenance, and the pattern of economic benefit.
Disclosure through the cash-flow statement removes the risk that depreciation estimates mislead investors.
Why it mattersDisclosure lets sophisticated investors adjust, but GAAP earnings still affect headlines, models, compensation, and multiples. A visible estimate can still be too optimistic.
Alphabet depreciates servers and network equipment generally over six years and bases useful-life estimates on asset performance, expected technological advances, and deployment plans.
CheckAlphabet's 2025 Form 10-K states both the general six-year life and the inputs used to estimate technical-infrastructure lives.
The capital spending that depreciation allocates is visible in public financial statements rather than hidden from investors.
CheckSEC filings disclose property and equipment, depreciation expense, capital expenditures, and operating cash flow. Investors can therefore analyze cash investment and management's useful-life estimates separately.
Will generative AI grow Google Search or merely let Google cannibalize it on its own terms?
Original point: Gemini lets Google cannibalize Search itself instead of surrendering that disruption to an outside model provider, but the transition remains a cannibalization problem.
What everyone argued
Chamath Palihapitiya
Chamath says Google has defended Search brilliantly and now controls the transition, but generative answers still force it to cannibalize its own legacy economics. Distribution through browsers, phones, and operating systems is Google's advantage as the market sorts into chat, enterprise, science, and other use cases.
Jason Calacanis
Jason takes the explicit other side: AI improves ad targeting and expands the number of searches, so total Search revenue can rise even if revenue per query falls. He predicts Google will defend and grow the franchise while OpenAI loses share to Google, Anthropic, Grok, and open models.
Winner circle
Jason wins the near-term round. Google did not just slow an outside attacker; Search revenue accelerated while AI features expanded query activity. Chamath keeps an important margin caveat, because public revenue does not reveal the economics of each AI answer, but the evidence through Q1 2026 favors growth over absolute cannibalization.
Commentary
Chamath Palihapitiya
Chamath has the better strategic frame: self-cannibalization can be a victory. He needed to separate revenue, gross profit, and query share, because a growing top line can hide weaker unit economics.
Assumptions and fact checks
AI answers will materially cannibalize Google's legacy revenue per search even if total Search revenue grows.
Why it mattersAI answers can reduce clicks and change ad inventory, but new query classes and new ad formats can offset that pressure. Public segment revenue does not reveal per-AI-query economics.
Owning Android, Chrome, and major distribution surfaces gives Google a durable advantage in consumer AI.
Why it mattersDefault placement and integrated workflows lower acquisition friction, though regulation and strong standalone assistants can constrain the advantage.
Google trained Gemini 3 on its own TPU infrastructure rather than NVIDIA GPUs.
CheckGoogle's Gemini 3 Pro model card identifies TPUs as the training hardware and JAX/ML Pathways as the software stack.
Jason Calacanis
Jason wins the narrow Search question with a measurable mechanism and subsequent revenue evidence. His OpenAI short is spicy but unnecessary; it adds a second forecast without helping prove Google's unit economics.
Assumptions and fact checks
Higher query volume and targeting gains will continue to outweigh lower monetization and higher serving cost per AI-assisted search.
Why it mattersReported revenue and query growth support the mechanism so far. The rating is not a permanent forecast: competitive pressure and inference cost could still change the balance.
Google's Search resilience makes OpenAI the clear short in an AI pair trade.
Why it mattersGoogle's success does not mechanically determine OpenAI's valuation or future revenue. Multiple providers can grow as the market expands, and private-company valuation adds separate timing and liquidity risks.
Google Search revenue and query activity were growing as generative-AI features rolled out.
CheckAlphabet reported Q3 2025 Search and other revenue of $56.6 billion, up 15%, and said AI experiences increased overall and commercial queries. Its Q1 2026 10-Q later reported $60.399 billion versus $50.702 billion a year earlier.
Chamath lands the cleanest correction: Jason's wording moved from 'USDT was used' toward 'Tether was involved.' He would have made a much stronger trust case by demanding the completed audit and controls evidence instead of treating a persuasive founder dinner as diligence.