The Waiver, the Window, and the Clock

On Thursday the White House published an executive order that placed a 500 percent tariff on Russian goods and a 100 percent tariff on the five largest buyers of Russian crude. Section 4 of the same order waives both, using a release valve Congress had written into the statute three weeks earlier. The tariff is law. The waiver is also law. When I pulled the price feed behind my own dashboard on Saturday morning, WTI crude sat at 101.84 dollars and Brent at 109.69, up about four percent on the quoted change, with the tariff scheduled to bite on Sunday.

That shape, a rule with an escape hatch, priced as though the hatch might be used and might not, is not unique to trade policy this week. It sits underneath a 35 billion dollar funding gap at the largest listed AI compute landlord, and underneath six million bitcoin whose public keys have been visible on-chain since 2009. A waiver, a window, and a clock. Three promises, each payable at somebody else's discretion, each repriced in the past five days.

Here is what the mechanisms actually are, what the numbers do once you check them, and where the honest counter-argument lives.

The Waiver#

The Graham Act, H.R. 5334, became law on 18 September, named for the late Senator Lindsey Graham, who died in July. It passed the Senate 86 to 11 and the House 262 to 159. The same statute extends the Iran Sanctions Act of 1996 through 2031. The implementing order arrived on 16 October. New purchases of Russian crude and gas on or after 18 October are what the 100 percent secondary tariff is aimed at, and the financial institution sanctions covering Sberbank, VTB and Gazprombank are now statutory under 22 U.S.C. 10802, with the relevant date 19 October.

Then comes the part the vote tallies do not capture. Section 115 of the Act lets the president waive any sanction, restriction or duty as long as he certifies to Congress in writing that the waiver serves the national interest. Section 4 of the executive order exercises that authority on day one, and the order states that both tariffs remain subject to the exceptions in Section 114 and to the waivers in Section 4. A Congress that voted 86 to 11 for mandatory pressure built the off-ramp into the same document, and the president drove through it immediately.

The enforcement machinery now sits with the USTR, which must track post-18 October purchases together with the State, Treasury and Energy departments and report recommendations back to the president. So the secondary tariff bites only if an agency determination is made and acted on. That is a promise to enforce, contingent on an internal memo.

Crude has already answered, at least in price. My dashboard's CNBC feed showed WTI at 101.84 and Brent at 109.69, and the move was upward on the quoted change. If enforcement is real, the rerouting friction alone lifts the cost of energy in a debt-saturated economy, and the Fed inherits a choice between tightening into a supply shock and easing into an inflation impulse. Both of those paths make the financing arithmetic in the next section worse. If enforcement is posture, the tariff is a negotiating chip and the barrel keeps flowing through the same ports under new paperwork.

I am not going to pretend this is unambiguous. The full scope of Section 4 requires the complete order text to settle, and whether the mandatory bank sanctions fall inside or outside the waiver is unresolved on the documents available. What is unambiguous is the structure: the headline rate is real, the legal exit is real, and the market has to price a probability rather than a rule.

The Window#

CoreWeave's three co-founders have sold 3.3 billion dollars of stock since the company's March 2025 IPO, according to Bloomberg's analysis of SEC Form 4 filings. Of that, 2.4 billion moved in 2026 alone. Michael Intrator sold more than 907 million dollars this year. Brannin McBee's total since listing stands near 1.06 billion, which Bloomberg puts at 40.4 percent of his undiluted stake. Brian Venturo moved more than 1.2 billion, over 35 percent of his original holding. Another 488 million dollars cleared in pre-IPO tenders. The lockup expired in August 2025.

The company's defence is that all of it ran through prearranged 10b5-1 trading plans. It notes that the founders remain among the top six shareholders, with roughly 20 percent combined. That is true as far as it goes. What the 10b5-1 defence cannot tell you is when those plans were set, and what the people who wrote them knew about the credit curve at that moment.

The credit curve is the actual story. CoreWeave's 9.625 percent unsecured notes due 2032 traded below 80 cents on the dollar at yields above 15 percent, per Bloomberg's reading of TRACE pricing. The five-year credit default swap spread hit an intraday record of 1,194.5 basis points, per ICE Data Services, a level S&P Global Market Intelligence's Gavan Nolan described to CNBC as generally a sign of credit distress.

Those two facts are the same fact. Do the plain bond arithmetic on that 9.625 percent coupon with six years left and it prices at 79.66 dollars at a 15 percent yield. A price below 80 and a yield above 15 percent therefore describe one market, not two. The swap market then converts the same fear into a probability. Run the standard approximation on that spread and the implied five-year default probability comes to 63.0 percent at the conventional 40 percent recovery assumption. Push recovery down to 20 percent and it falls to 52.6 percent; at 50 percent it rises to 69.7 percent. The widely repeated figure of more than 60 percent therefore assumes recovery of roughly 40 cents on the dollar, which is the convention for an unsecured claim in this sector. My numbers, their spread.

Now the funding side. CoreWeave carries about 35 billion dollars of debt, and roughly 51 billion dollars of total obligations including lease liabilities on its Q2 2026 filing. It raised full-year capital expenditure guidance to between 35 and 39 billion dollars. Deutsche Bank's note on the gap is the cleanest version of the problem: near-term operating cash flow of roughly 10 billion dollars against required capex of about 45 billion, a shortfall near 35 billion. If half of that were raised at current distressed rates of 15 percent, the added interest would run to 2.63 billion dollars a year by my arithmetic, against reported consensus losses already in the billions.

Cory Klippsten's numbers put the same squeeze on the first half of this year. CoreWeave spent 14 billion dollars on equipment and infrastructure against less than 4 billion in operating cash. Roughly three quarters of the build was funded from outside. PIMCO's September consensus forecast put next year's combined hyperscaler cash shortfall, after capital spending, at 87 billion dollars.

There is a structural detail that turns this from one company's problem into a market's problem. CoreWeave is simultaneously the borrower on its own unsecured junk bonds and the tenant whose lease payments backstop a chunk of the junk-rated datacenter bonds issued by other companies. A default or a lease renegotiation does not stay inside CRWV equity. It travels upstream into the bonds those lease payments were meant to service. The exposure is circular rather than linear, and the neocloud refinancing wall lands between 2026 and 2028.

The cost side keeps climbing while the revenue side has one number worth watching. The Pomp Letter's Friday charts, drawing on Sage Road Research, show switchgear prices up about 60 percent since the end of 2021. Transformers, cooling equipment and generators are each up around 30 percent, and general construction materials up about 10 percent. Datacenter memory costs more than five times what it did a year ago. Only 492 billion dollars of a 2.3 trillion dollar datacenter pipeline is under construction. The remaining 1.8 trillion gets built at today's prices. Against that, OpenAI has held roughly 10 billion dollars of revenue per gigawatt since 2023, by the chart in the same letter. Anthropic's July run rate of more than 65 billion implies somewhere between 13 and 32.5 billion per gigawatt, depending on whether it runs two gigawatts or five.

One more chart deserves a second look. Open-source models handle 78 percent of the tokens through one major AI gateway, while closed models from OpenAI and Anthropic handle 22 percent. But seven of the top nine models by spend are closed, and the two open models in that top nine, Kimi K3 and DeepSeek, account for just 14 percent of dollars. Volume is migrating to cheap models; money is not. Cheaper tokens mean more tokens, and every token still consumes chips, memory and power. That is the quiet bull case for the supply chain, and it is also the reason the payer does not need the application layer to earn a return.

The Clock#

On 15 October, in Science, Google Quantum AI reported that its 105-qubit Willow processor can prepare magic states with roughly 40 times fewer errors than uncultivated states. The experiment used at most 54 of the 105 qubits. It reached a T-state logical fidelity of 99.99 percent among the roughly 8 percent of preparation attempts that passed every quality check. Emma Rosenfeld and Craig Gidney co-authored the accompanying post. Rosenfeld had already described the scale of the problem: magic state generation can eat up to 95 percent of a fault-tolerant machine's total capacity.

That is why this result matters more than its headline suggests. Magic states are the resource that makes non-Clifford gates possible, and under traditional distillation they consume 90 to 95 percent of the qubit budget. Cultivation attacks the most expensive line item in every blueprint. The theory was published in 2024 by Gidney, Noah Shutty and Cody Jones. October 2026 is the first hardware demonstration. Moving from paper to silicon is the kind of milestone that compresses projections, and it is still not a break.

The distance left is measurable. Cracking RSA-2048 or bitcoin's ECDSA keys needs magic state error rates near one in ten billion. Willow achieved roughly one in ten thousand. That is a factor of one million, by my own division of the two published figures. Errors shrink with the cube of the chip's physical noise. Halving hardware noise therefore takes magic state errors to about 1.25 in 100,000 in my reading of that scaling, close to the roughly 2 in 100,000 Google states. Each generation compounds.

The other number that moved this year is the qubit requirement itself. A March 2026 paper from Google Quantum AI, co-authored with researchers from the Ethereum Foundation and Stanford, concluded that bitcoin's secp256k1 curve could be broken with fewer than 500,000 physical qubits. That is roughly one twentieth of the earlier consensus estimate.

So count the exposed supply. I ran this against my own node this morning: block 971,827, which by protocol issuance is 20,099,462.5 bitcoin in existence. The twenty-millionth coin was mined at block 940,000, according to that node's header, which dates it 9 March 2026. Glassnode's count of coins in addresses whose public key has been permanently broadcast is roughly 6.04 million, which is 30.05 percent of issued supply on my division. Europol's 7 October warning used a wider figure of about 6.9 million, or 34.3 percent. About 1.1 million of those coins, 5.47 percent of supply, sit in Satoshi-era pay-to-public-key outputs from 2009, where the key was hardcoded on-chain from the moment the block was mined.

The exposure classes are not equal. A pay-to-public-key output gives an attacker the public key without any action by the owner. An address that has ever sent a transaction has broadcast its key permanently, in the signature it spent. A modern address that has never spent has not. That last category is the only one with time to spare, and only until a cryptographically relevant machine exists.

Which brings the argument to the part that should worry people who care about property rights more than about timelines. Satoshi-era coins cannot be migrated. The keys are lost. Any proposal to protect them is a proposal to render specific unspent outputs unspendable by consensus rule, based on address type rather than any demonstrated theft. That is a freeze, and it would set the precedent that the network can override ownership on an external threat assessment, with no answer to who controls the resulting escrow. The Quantum Alliance, formed by nine companies on 15 October, is not the counterweight here. Its roster runs Quantus, American Fortress, Intmax, Keystone, Midnight Foundation, Postquant Labs, Abelian, The Vault and Hacken, which is mostly not bitcoin. The Alliance's Quantum Doom Clock points at March 2028, while independent researchers place the more likely window in the early-to-mid 2030s. The bitcoin-native work is BIP-360, a pay-to-merkle-root address proposal whose July 2026 revision added Python examples and test vectors, and which sits in Bitcoin Optech's October discussions on post-quantum output design, wallet compatibility and signature cost. No press release required.

The Same Trade#

Pull the three apart and the same instrument is left on the table. In each case, what changed price is not an asset but the probability that a promise gets performed. A tariff whose enforcement rests on a USTR determination. A take-or-pay contract whose counterparty must keep paying whether it uses the capacity or not. A key migration that only works if the owner still holds the key. Every one of those is a claim on a future discretionary act, and the discount rate applied to discretion moved this week.

The banks demonstrate the cleanest version, because the promise there is safekeeping. Bank of America earned 4.61 percent on the 363 billion dollars it held in fed funds and reverse repos last quarter. More than half a trillion dollars of its customers' deposits paid nothing at all. That gap is how a single quarter books 16.4 billion dollars of net interest income. Separately it paid a 1.2 billion dollar penalty for seven years of willful Bank Secrecy Act violations, after reporting 8.1 billion dollars of profit in the same quarter. That ratio is the price of the promise.

The same week, the Office of the Comptroller of the Currency published its debanking review of the nine largest banks. All nine had policies restricting access for lawful businesses, and bitcoin and crypto companies were grouped with coal, firearms and tobacco, with the restrictions often attributed to financial crime considerations. The regulator wrote that money laundering compliance could be used as a pretext for other grounds. The lobby that benefits from that free float spent the year fighting a dollar product that would pass part of the Treasury yield back to depositors. The American Bankers Association and four other groups declared that the correct policy goal was prohibiting the payment of yield on stablecoins. The Clarity Act failed a procedural vote 49 to 50 on 15 September, and eight banking groups asked Congress for targeted changes to stablecoin yield policy the same day.

Set that beside Nik Bhatia's observation at The Bitcoin Layer, which runs the other way and is worth taking seriously. He argues that a Treasury yielding above five percent recruits its own buyers. The level offers positive carry against SOFR financing, downside protection from the coupon, and positive convexity from sitting on that part of the price-yield curve. By that reading, the same rate level that makes a leveraged AI landlord unviable is what makes the sovereign's own paper self-supporting. In my dashboard's reading this morning, the US ten-year sat at 5.081 percent and the British gilt at 5.3771. The German bund traded at 3.4206 and the Japanese ten-year at 2.937. None of those levels is a crisis. All of them are a repricing.

The Case Against My Own Reading#

The bear case on AI financing has a specific way to be wrong, and it is not a vague hope. CoreWeave raised more than 4 billion dollars in convertible bonds in September, which is not the behaviour of a company shut out of markets. Its largest customers have cash flows and financing options that no neocloud has. Deutsche Bank still carries a buy rating and a 150 dollar price target. Jensen Huang said in Stockholm on 15 October that CoreWeave's cost of capital is actually decreasing, which is a claim the swap market disputes, but he has better access to the loan book than I do. If the 35 billion dollar gap closes below ten percent, the founder sales of 2025 and 2026 will look like ordinary concentration management, and the CDS print will read as a spike rather than a trend.

The quantum timeline has a specific way to be wrong too. TFTC frames the falsifiable question well: whether cultivation's error gains survive the move from 54 qubits to the hundreds of thousands a cryptographically relevant machine needs. An 8 percent acceptance rate is tolerable in a laboratory and potentially useless at industrial volume. If the 40-fold reduction does not replicate at scale, the "millions of qubits" comfort returns, and bitcoin's deliberate consensus process stops looking slow.

The tariff waiver has the simplest out of all. If the order is leverage rather than policy, oil retraces and the Fed gets its choice back. The credit math then stops deteriorating for reasons that have nothing to do with any of this.

What to Watch#

Four dates carry the argument, each checkable rather than interpretive. Sunday 18 October, when purchases of Russian crude fall under the secondary tariff and the USTR's tracking obligation begins. Monday 19 October, the statutory date for action against Sberbank, VTB and Gazprombank, where the question is whether the waiver reaches them. The next CoreWeave funding print, where the coupon is the vote on whether capital is still available below ten percent. And the next BIP-360 revision, which is the only quantum answer that does not require anyone's permission.

The pattern across all three stories is that the market has started charging for discretion. A statute that waives itself, a founder exit that closed before the lenders' did, and a key that has been public since 2009 are three assets with one shared property. Their value depends on a decision someone has not made yet. The consequence is that prices are now attached to that decision point, which is the one input a spreadsheet cannot model and a bond market can, given enough trades. The bill for the promises is being marked to market one discount rate at a time, and this week the market did the auditing that no press release could.

Sources#

• TFTC / Marty Bent — Trump EO Activates 500% Russia Tariffs While Waiving Its Own Sanctions
• TFTC / Marty Bent — Google Willow Cuts Quantum Error Rate 40x, Tightening Bitcoin's Post-Quantum Clock
• TFTC / Marty Bent — Nine Companies Launch Quantum Alliance as 6 Million BTC Face Exposure Risk
• TFTC / Marty Bent — CoreWeave Founders Dumped $3.3B in Stock as Bonds Cratered
• TFTC / Marty Bent — The Banks Above the Law
• Cory Klippsten — What Happens to the AI Build-Out When the Next Loan Gets Harder to Obtain?
• The Pomp Letter — Five Friday Charts: Where AI money really goes
• The Bitcoin Layer / Nik Bhatia — What If US Treasury Yields Go Back To 3%?
• Checkonchain / James Check — Divergent Treasury Strategies
• CNBC — U.S. 10 Year Treasury yield quote
• CNBC — WTI Crude Oil quote
• CNBC — Brent Crude Oil quote
• Bitcoin Core RPC — block height, issuance and block 940,000 header (own node, 17 October 2026)
• Europol — post-quantum migration warning on exposed crypto addresses
• Bitcoin Optech — post-quantum output design discussion
• BIP-360 — Pay to Merkle Root addresses
• The Quantum Insider — Quantum Alliance launch and Quantum Doom Clock
• Bloomberg — CoreWeave founder stock sales analysis of SEC Form 4 filings
• ICE Data Services — credit default swap pricing
• The Graham Act, H.R. 5334 — tariff, waiver and Iran sanctions provisions