DEV Community: AI & Markets

What Happens Next: The AI Bill Has a Calendar

AI & Markets — Fri, 23 Oct 2026 23:53:30 +0000

AI & Markets - David Groch

Two markets are now pricing two different revolutions.

The equity market is pricing demand, and on the evidence it is right. South Korea's September exports rose 83.5% year over year to a record $120.9 billion, with semiconductor shipments up 263% to more than $60 billion in a single month. That is not a narrative. That is customs data. Somebody is buying every chip Asia can make.

The credit market is pricing the bill, and on the evidence it is also right. Five-year credit default swaps hit records across the AI complex this month: Oracle at 322 basis points, up from 144 in January; Broadcom at 136; Nvidia at 101; Meta at 108; SpaceX at 226; CoreWeave near 1,090, which is distressed territory by any definition. Converted into the market's own arithmetic, Oracle's spread implies roughly a 23% five-year default probability, Broadcom's about 12%, Nvidia's about 8%. AI-linked companies have issued an estimated $466 billion of debt in 2026, against $101 billion in the same period of 2025. Weekly trading in CDS on the seven biggest AI names went from under $2 billion in July to over $9 billion in September.

One honest caveat before the calendar: an implied default probability is not a forecast. Part of the widening is mechanical, because equity managers buy CDS as a cheap hedge and hedging demand widens spreads on its own. Oracle is not going bankrupt at a 23% rate in any world a credit analyst would sign. What the spreads actually say is narrower and more useful: the price of insuring the buildout has multiplied, the marginal buyer of AI debt now demands to be paid for duration and doubt, and the next round of funding will not be priced like the last one. That is the story. And unlike most market stories, this one has dates.

Next week: the pincer

Monday, October 26. Celestica reports after the close, with its investor day on Tuesday. This is the margin canary. Celestica is a contract manufacturer with thin EMS margins and enormous hyperscaler exposure, and the one serious bear case on the name is simple: when AI hardware margins compress, they compress for the assembler first. The Street is unanimously bullish (average target near $474 against a $262 stock), which means the print has to be clean, not good. A margin wobble here is the first company-level confirmation that the buildout's economics are squeezing the supply chain. The same morning, the first big SpaceX lockup tranche (up to 328 million shares) becomes tradable, a week after the stock rose 8% on earnings and fell 9% on profit-taking and unlock fears. Supply meets a story stock.

Wednesday, October 28. The week's hinge, and it is a pincer. At 2:00 p.m. ET the FOMC decides. The Reuters poll of 100 economists is as close to unanimous as polls get: 94 expect a hold at 3.75% to 4.00%, and all 100 expect a hike in December. The risk is not the decision, it is the language, because Chair Warsh has dropped forward guidance and publicly distanced himself from the committee's own projections. Two hours later, Microsoft and Alphabet report. Their earnings are no longer really about earnings. Meta already showed the template: beat on revenue, guide above consensus, and lose 7% anyway because the capex line moved. What Microsoft and Alphabet say about 2027 spending is the referendum on whether Meta's repricing was one company's problem or the sector's.

Thursday, October 29. September PCE and the first Q3 GDP print land before the open (consensus: core PCE 3.1% year over year, headline 3.6%), and Amazon reports after the close. A hot core print with a December hike already unanimous is the stagflation-lite combination: it validates the CDS move and pushes the terminal-rate debate from one more hike to two. A soft print is the only scheduled event all week that can actually relieve the credit market.

Next month: the report cards

November 3 brings the US midterms, which matter here for one reason: energy and industrial policy are now market variables, and oil above $110 with the Strategic Petroleum Reserve at its lowest since 1982 (282.7 million barrels) is a political fact as much as an economic one.

November 2 to 12 is the cluster that decides whether the CDS story stays a hedge or becomes a verdict. The leveraged end of the buildout reports in a ten-day window: Cipher Mining on November 2, Core Scientific on November 5, CoreWeave around November 9, Nebius around November 10, TeraWulf in the same stretch. Read these reports for one line only: the funding line. Bank of America's preview of CoreWeave is the template of the whole moment. The bank rates the stock a buy with a $140 target, forecasts Q3 revenue of $3.66 billion (up 168%), cites a $129 billion backlog, and then estimates the company needs roughly $20 billion of additional funding, with $12.5 billion of capex in the quarter alone. Even the bull case is now written as a financing story. Backlog is not the question anymore. The coupon is.

November 18 is Nvidia, the last and largest report of the season, and throughout the month Meta's $25 billion New Mexico judgment accrues interest at 15% a year, roughly $10 million a day, which is its own kind of calendar: every week of appeal is a week the settlement price argument changes.

Just past the horizon, December holds the two dates that could close the loop: the FOMC meeting where 100 of 100 economists expect the hike, and December 9, when Micron's $35 billion buyback authorization becomes usable. The largest memory maker becoming a structural buyer of its own stock is the equity market's best scheduled answer to the credit market's question.

Three ways this resolves

The bill gets paid quietly. Microsoft, Alphabet and Amazon hold or trim their capex guides to something the cash flow can carry, the neoclouds fund at tolerable rates, and CDS spreads tighten into year-end. Signpost: Oracle CDS back under 250 basis points and a CoreWeave funding deal at single-digit rates. In this world the October repricing was a tantrum.

The slow bleed. Guides go up again, the Fed hikes in December with two more on the table, and equities de-rate gradually toward where credit already is. No crash, no capitulation, just a multiple that shrinks while the spending does not. Signpost: bank stocks, already on their longest losing streak since 2008 (six straight weeks for the KBW index), keep sliding while the indexes hold up on a narrowing group of names.

Something breaks at the weak end. A funding deal fails, or prices at the mid-teens yields where CoreWeave's bonds already flirt, and the question stops being about one neocloud's balance sheet and becomes about who else financed the same way. Signpost: any postponement or pulled deal in the November cluster, or CoreWeave CDS pushing through 1,200 basis points. This is the only scenario where the Asia story and the credit story collide, because the chip demand is real, and real demand financed badly is exactly how the last buildout ended: the fiber got lit, the carriers did not survive it, and the cables outlived the companies that laid them.

The demand is not the debate. The price of money is. And for once, the market has published the schedule on which it will argue: October 28 for rates, October 28 and 29 for capex, November 2 to 12 for the leveraged end, December for the bill itself. Watch the calendar, not the commentary.

Facts and sources (checked October 23, 2026): CDS levels and implied probabilities, Morgan Stanley research via the Financial Times and Financial Gurkha (October 19-20), Mott Capital via TalkMarkets (October 19). AI-linked issuance $466B and CDS weekly volume, Morgan Stanley via FT. Korea September exports, Reuters and MOTIE customs data (October 1). FOMC expectations, Reuters poll of 100 economists (October 16-20). PCE and GDP consensus, Morningstar week-ahead (October 23). Earnings dates: Microsoft and Alphabet October 28, Amazon October 29 (company notices via Admiral Markets, Reuters, Zacks); CoreWeave around November 9 (Zacks estimate, company date not formally announced); Nebius around November 10 (StockAnalysis estimate); Nvidia November 18 (TradeStation calendar, estimated); Core Scientific November 5 (company). CoreWeave funding estimates and $140 target, Bank of America (October 22). SPR level 282.7M barrels, DOE data via Reuters (October 19). KBW Bank Index six-week losing streak, Bloomberg via TradingView (October 20). Celestica consensus and targets, Seeking Alpha. Implied default probabilities are model outputs that depend on recovery-rate assumptions; they are not forecasts.

Inflation came back. The market yawned.

AI & Markets — Fri, 16 Oct 2026 15:48:05 +0000

September CPI: 3.7% year over year, 0.6% in a single month. Gasoline did most of the damage, up 9.1% in September and 34% on the year. Then PPI followed: up 1.3% on the month, 6.3% on the year, the hottest wholesale print since March 2022. The pipeline is full, and it points one direction.

This is now 67 straight months of US inflation above the Fed's 2% target. The last sub-2% reading was in 2021. A dollar kept under the mattress for those five years buys about 22% less today. That is the quiet tax nobody votes on.

And the market's response has been a shrug. Not because the numbers are fine, but because the market is trading something else entirely: the calendar.

A Truth Social post set a date. No strikes on Iran before the midterms, talks "productive," blockade still in full force. Brent slipped two dollars and the tape called it a relief rally. Brent is still $105. Diesel costs more than it did before the celebrated "deal" with Putin. Ten ships have been hit near Hormuz since Sunday. Nothing was resolved. Everything was rescheduled, and the market priced the reschedule.

This is the same trade all year. Volatility arrives, looks at the calendar, and leaves. The half-life of a scare in 2026 is about eighteen hours. It works until the date expires. The Iran pledge expires November 3.

Meanwhile the other repricing is already underway, and it is not about geopolitics. It is about who pays for the AI buildout. CoreWeave bonds yield over 15%. The KOSPI lost 20% in seven trading sessions. Semis just logged eight straight losing days. Gartner is warning that trillions in AI infrastructure spending could face impairment by 2030. Anthropic posted $16.8B in quarterly revenue with growth slowing, and is rushing an IPO next month while the window is open.

Here is the uncomfortable part. Most stocks have been living in a bear market for months. Money keeps flowing into the AI trade anyway, ignoring every red light, with hedge fund concentration in the top ten tech names at records. The market is long postponement and short the bill. Inflation is the bill, and it is already in the numbers: 3.7% and rising, 6.3% at wholesale, $105 oil, record gas prices.

Bonds noticed before stocks did. The 10-year fell this week even on hot inflation prints, because money ran for safety during the equity selloff. When bond buyers start ignoring inflation to hide from stocks, the rotation is real.

Watch two things. The Strait, where the physical barrels move, and the calendar, where the postponements expire. Everything else is noise priced as signal.

The market doesn't need peace. It needs a calendar.

AI & Markets — Fri, 09 Oct 2026 18:11:32 +0000

Originally published on AI & Markets: https://aiandmarkets.substack.com/p/the-market-doesnt-need-peace-it-needs

Thursday opened with Brent above $105. By Friday morning it was under $103 and the Dow was up 400 points. In between, exactly one thing changed: a Truth Social post set a date.

This is the actual market driver of 2026, and it is not complicated. Nobody is pricing peace. They are pricing the postponement. "Not this week, not this month" is the new good news, and the tape is buying it in size.

Consider the oil. A few weeks ago, WTI at $90 was a shock, the number that launched a thousand gas-station memes. This week WTI is sitting near $91 and the market treats it like the all clear. Brent is at $103, down from $105.88 on Thursday, and that two-dollar slide counts as a relief rally. The price did not get better. The reference point moved. The market has re-anchored, and it re-anchored fast.

That is the tell. Anchoring is supposed to be slow. It took years for $100 oil to feel normal in 2022. This time it took weeks, because the war did the anchoring for it. When the fear is big enough, anything that is not the worst case reads as a bargain. Ninety dollars is not cheap. It is just cheaper than the thing everyone was imagining on Thursday morning.

Now the mechanism, because the mechanism is the trade. On Wednesday, The Atlantic reported the White House had asked the Pentagon for limited Iran strike options usable before the November 3 midterms. The New York Times added that a three-day campaign was being discussed. Brent spiked more than 4% on Thursday. Then the post landed: the US will not attack Iran "at any time prior to the Midterm Elections," the blockade stays, talks are "productive." Oil gave back the gains. Stocks cut their losses. Heading into Friday's close, the S&P is up about 1.1% on the week, the Nasdaq up 0.6%, the Dow up 0.9%. Volatility arrived, looked at the calendar, and left.

Read the fine print and it gets stranger. The pledge leaves post-midterm strikes explicitly on the table. Trump told Time earlier this month that ramped-up strikes were "possible" after the election. The Pentagon just moved a third carrier group and 10,000 troops into the region. On Thursday, the same day as the peace post, the Treasury sanctioned 17 vessels in Iran's shadow fleet. The talks are "productive" while the blockade stays "in full force and effect." Nothing was resolved. Everything was rescheduled. The market priced the reschedule.

This is not a new trick. It is the oldest pattern in the book, and 2026 keeps running it back. Thursday's other scare was a Financial Times report that OpenAI's annualized revenue is $50 billion, not the $68 billion floating around outside estimates. The Nasdaq fell 1.3%, its worst day since mid-August, the SOX dropped 3.5%. By Friday morning analysts were calling it an accounting difference, futures were up 0.8%, and the tape was green. The half-life of a scare this year is roughly eighteen hours. Every test of investor appetite gets absorbed, and the rally extends, because the appetite is not really about the news. It is about whether there is anyone left to sell. There usually isn't, for long.

Even the bond market's screaming does not dent it. The 10-year is at 5.24 to 5.29%, up almost 120 basis points this year, having touched a multi-decade high of 5.364% this week, the highest since 2002. Pimco's chief investment officer told the FT that 6% is "feasible" near-term and that 5.5% would bring "decent weakness in risk markets." Meanwhile spot bitcoin ETFs bled $731 million over two sessions, October is net negative, and the US government moved $1.54 billion in seized bitcoin into Coinbase Prime custody over three days. Each of these is, on paper, a reason to hide. The market's answer is the same shrug: not this week, not this month.

So the real question is what breaks the spell. Two things, and they are the same thing. First, a genuine supply disruption. Ritterbusch's note Friday said Trump's comment put crude back on the defensive "amid what is likely to be a lull in attacks through the Strait of Hormuz." A lull. The physical market is telling a different story than the futures: supertanker freight rates are at records, around $82 million for a US Gulf-to-Japan voyage, and the Brent-WTI spread is $12.42, which is the Hormuz risk premium made visible. A tanker was hit off Qatar on Wednesday. Hurricane Isaias, now Category 3, has Gulf platforms evacuated and landfall expected this weekend. The paper barrel is calm. The physical barrel is not.

Second, the calendar runs out. The pledge expires November 3. Everything that was deferred is still on the table: the strikes, the campaign, the three carriers. The market is buying a date, not an outcome, and dates expire. Postponement is a trade with a fixed expiry. Nobody is pricing what comes after it, because there is no trade in "we'll see."

Here is the part the tape will not say out loud. While the market celebrates the postponement, the real economy is paying the tax the market doesn't feel. The national average for a gallon of gas is $4.36, per AAA, the highest it has ever been in October. September set the monthly record at $4.33, fifty cents above the old record. California drivers are paying $6.39. The September EIA average is up 37.6% from a year ago. That is the war, still on, priced into every commute, while the market throws a party because the next headline got moved to December. The disconnect is the story. The war was the shock. Now the shock is the baseline, and the market only needs the baseline to hold.

Strip it down: markets want to grow, and growth only needs one fuel, uncertainty that is shrinking. Not gone. Shrinking. A war that is cooling is better than a war that is over, because a war that is over demands a new narrative, and a war that is cooling just demands patience. The market has infinite patience for anything labeled "later." It has none for "now."

That works until later becomes now. Watch the two things that actually move physical barrels: the Strait and the calendar. Everything else is noise priced as signal. The relief rally is priced on a postponement. Postponements end. When this one does, the market will discover it bought a date, and the date will be past, and $90 oil will be a shock again.

Private marks meet the public tape

AI & Markets — Thu, 08 Oct 2026 15:55:30 +0000

The private market has had its say. Anthropic went from an $18 billion valuation in early 2024 to $965 billion this May, doubling roughly every three months for fourteen months. Now the company is headed for the public market, with reports pointing to a roadshow the week of November 9 and a listing before Thanksgiving. The number being discussed is $2 trillion, with up to $100 billion raised. If it happens, it would be the largest IPO in history, eclipsing the record SpaceX set in June.

This is the moment private optimism meets public pricing. Here is what the public is being asked to buy.

Start with the growth, because it is real. Anthropic's recognized revenue jumped twelvefold in 2025 to $4.6 billion, per its leaked S-1 filing. The first quarter of 2026 alone did $4.73 billion. The annualized run rate crossed $65 billion by July, and the company reports positive adjusted operating income in the second quarter. Some reports now put the pace above $100 billion, though those are unverified whispers rather than audited numbers. Roughly 80 percent of revenue comes from enterprise customers: more than 300,000 businesses, over 1,000 spending more than a million dollars a year, eight of the Fortune 10. This is not consumer hype. It is committed enterprise spend, per token, on a product that demonstrably works.

Now the other ledger. Anthropic posted an operating loss of about $8 billion in 2025. The headline net loss was around $42 billion, though roughly $34 billion of that was non-cash accounting charges tied to financing instruments rather than operating cash burn. And then there is the stack of commitments: at least $518 billion of infrastructure obligations over roughly the next decade, per Reuters, with about 80 percent noncancelable or payable whether Anthropic uses the capacity or not. Hyperscaler commitments with Alphabet, Amazon, and Microsoft total about $252.5 billion, with minimum spend floors. Broadcom-linked equipment leases account for another $161 billion.

Read that again. A $100 billion IPO, the largest in history, covers roughly one fifth of the mostly-fixed commitments already signed. The raise is not the funding event. It is one tap among several the company will have to keep turning.

This is heavy-industry economics wearing software multiples. At a $2 trillion valuation against a $65 billion run rate, the multiple is about 31 times revenue. Against 2025 recognized revenue, it is over 400 times. On a business that lost $8 billion on operations last year and owes half a trillion dollars in mostly-fixed payments. The bull case is that inference costs keep falling while usage compounds faster, so revenue per dollar of compute improves over time. The bear case is simpler: cheap AI makes customers consume vastly more tokens, so the infrastructure bill keeps growing with the revenue.

There is a recent precedent for what happens when record private optimism meets the public tape. SpaceX went public in June at $1.77 trillion, the largest IPO ever. Day one was euphoric: shares opened at $150 against a $135 offer, closed up 19 percent, and rallied past $225 in the days after, briefly touching a $2.1 trillion market cap. Then gravity arrived. Within weeks the stock traded below its IPO price. Day-one buyers who held are underwater. The skeptics had done the math in advance: Morningstar pegged fair value at $780 billion, less than half the IPO price. SpaceX had $18.7 billion in 2025 revenue and a GAAP net loss. The IPO priced it at 95 times revenue, 112 times at the debut-day peak. Euphoria is not a valuation method.

Anthropic's offering arrives into a harsher macro backdrop than SpaceX's did. The 10-year Treasury sits near 5.35 percent, a 24-year high, and the September FOMC minutes show most officials still see another hike this year. A 5.35 percent risk-free rate reprices everything that is not a proven cash machine. OpenAI has delayed its own IPO to 2027. Other planned listings have been pulled for poor sentiment. The tape is printing records into this IPO, but the bond market is quietly arguing with it.

Then there are the structural quirks worth knowing. Anthropic's Long-Term Benefit Trust holds special shares with governance powers, and founders are reportedly getting super-voting stock, so public shareholders will get economics, not control. The hyperscalers writing the biggest checks are also the suppliers owed the biggest commitments, which makes for complicated incentives. And the CEO who called for the industry to slow down over safety concerns is the same CEO asking the public for $100 billion. That tension will not resolve on listing day.

None of this is an argument that Anthropic is a bad company. It is arguably the best-positioned AI lab on earth, with real revenue compounding at a pace no company in American history has matched. The question is only the price. Private marks are set by mutual agreement among people who all benefit from the mark going up. Public marks are set by everyone, every second, including people with no reason to be kind. SpaceX taught that lesson three months ago. Anthropic is about to get its own exam, at $2 trillion, into the teeth of 5.35 percent yields.

The tape is printing records. The bonds are asking questions. November will settle which one was right.

Originally published on AI & Markets:

https://aiandmarkets.substack.com/p/private-marks-meet-the-public-tape

The AI buildout has a power problem

AI & Markets — Tue, 06 Oct 2026 04:02:19 +0000

The first piece in this series argued the AI buildout is a financing story: $700-730 billion in 2026 hyperscaler capex, up 70-80% from last year, and money is not the constraint. This one is about the constraint that actually is.

The industry is spending $700 billion a year to build data centers faster than anyone can power them.

The demand is not theoretical.

Lawrence Berkeley National Lab's August 2026 update puts US data-center electricity at 521-843 TWh by 2030, with a reference case of 649 TWh, or 11.8% of all US electricity. Goldman Sachs sees US data-center power demand going from 31 GW in 2025 to 66 GW in 2027, from 4.1% to 8.5% of peak summer demand. Moody's puts the bill at roughly $110 billion for 45 GW of new generation by 2030. Globally, the IEA has data-center consumption rising from ~415 TWh in 2024 to ~945 TWh by 2030.

Treat these reference cases as ceilings, not central estimates.

Big Tech is buying electrons. Some are real.

The deal flow of the past year splits cleanly into two buckets: contracted electrons with dates, and capacity options dressed as announcements.

In the first bucket: Microsoft's 20-year PPA with Constellation to restart Three Mile Island Unit 1, 835 MW, targeted for 2027. Amazon's $650 million for a campus co-located at the Susquehanna nuclear plant, 960 MW initially rising to 1.92 GW through 2042, wired directly behind the meter. Oracle's October deal for 125-250 MW from Wisconsin's Point Beach plant.

In the second bucket: Google's deal with Kairos Power for up to 500 MW, first 50 MW unit around 2030. Amazon's $700 million into X-energy for up to 12 SMRs. Meta's three agreements totaling up to 6.6 GW across TerraPower, Oklo, Vistra, and Constellation, with first power "as early as" 2030.

Add it up and Big Tech has contracted more than 10 GW of new nuclear capacity in the past year. But notice the structure: the restarts and co-locations are power. The SMR deals are development pipelines. Options, not electrons.

The SMR backlog race, ranked by paper quality.

Three public companies are competing to be the SMR story, and their backlogs tell you everything about how seriously to take each one.

Oklo has the biggest headline: 18.1 GW at end of 2025 (about 14 GW in 2026 reporting), though roughly 12 GW of that comes from a single customer under a largely nonbinding deal. But it also has two things neither rival can claim. First, the only Big Tech check actually written to an SMR developer: Meta's January deal for up to 1.2 GW at a Pike County, Ohio campus, where Meta prepays for power and funds early development. Google and Amazon signed future-power agreements; Meta put money down. Second, the fastest plausible path to first power: Oklo's Aurora unit targets 2028 on the DOE pilot track at Idaho National Lab, which can move quicker than the full NRC licensing queue everyone else is stuck in. Oklo also sells electricity rather than reactors, so its incentives point at operating plants, not winning design contests. It reported its first-ever revenue in Q2 2026: $1.2 million, from services, not power. Early, but real.

NuScale has the best paper of the three and the worst precedent. It's the only SMR with a full NRC design certification (January 2023), and its uprated 77 MWe module got design approval in May 2025. Its backlog is 6-7 GW, nearly all from one eastern-US utility customer, largely nonbinding. Its most advanced project, in Romania, now expects first commercial operation in 2033.

GE Vernova is the one actually pouring concrete. Its BWRX-300 is under construction at Darlington in Ontario since May 2025, with Unit 1 targeted for 2029. The NRC issued the first US construction permit for the design to TVA's Clinch River site in September 2026. Nuclear is under 3% of GE Vernova's sales; the SMR is an option, not the thesis.

The honest ranking: GE Vernova has the only project under construction. NuScale has the only certified design and no firm build. Oklo has the largest headline on the weakest paper, but also the only Big Tech check and the fastest plausible route to first power. And none of the three has delivered a commercial electron from a new-build SMR.

The skeptic's ledger.

Small modular reactors have already failed on cost once. NuScale's flagship Utah project with UAMPS was cancelled in November 2023 after the price went from $4.2 billion to $9.3 billion and the target price from $58 to $89/MWh, up 53%. Even DOE's $1.36 billion cost-share couldn't save it. Current SMR cost estimates run $100-180/MWh, and a European study puts average SMR capital cost at EUR 7,031/kW, about 41% above large reactors.

The bridge fuel is bottlenecked too. Gas turbine costs are up 195% since 2019. Industry output is 60-70 GW a year against 110+ GW of orders. GE Vernova's gas equipment backlog sits at 116 GW, with delivery slots offered into 2031. Operators are paying $25 million just to reserve a 2030 slot, and non-refundable reservation fees are now standard.

The fuel doesn't exist at scale. Centrus, the only US producer of the high-assay uranium advanced reactors need, has made less than 2 metric tons since 2019. Large volumes don't arrive before 2029.

US residential electricity rates are up 42% since 2019. PJM's December capacity auction cleared at the price cap and still came up short, with prices up 9x since 2023. Oracle's Wisconsin deal is at the center of a proposed $176 million rate hike fight.

When do the electrons actually arrive?

Three Mile Island restarts in 2027. Oklo targets 2028 for its first Aurora unit, company guidance, with the NRC license still pending. Darlington's first unit is targeted for 2029. TerraPower's Natrium, Kairos's first 50 MW unit, and the Meta-Oklo campus all point to 2030 "at best." Romania is 2033.

Read that list again. The only new nuclear electrons contracted for this decade come from restarting reactors built in the 1970s. Everything else is 2029-2033, at best, assuming nothing slips. Nuclear has never not slipped.

What fills the gap is natural gas (more than 30 of Moody's 45 GW), grid uprates, behind-the-meter projects (about 30% of new capacity is going on-site), aeroderivative turbines, and fuel cells. Fossil fuels still supply more than 60% of AI power today.

One more risk, on the demand side.

This week, four rival lab CEOs agreed for the first time that the frontier should slow down, and the infrastructure trade sold off the same day. It may be a press cycle. But file it: if the labs ever mean it, every 2030 demand forecast above is the bull case.

So where does it break?

Not a prediction. A watchlist. Whether Three Mile Island actually restarts in 2027 on schedule. Whether Oklo's NRC license application survives contact with the regulator that denied its 2020 application. Whether Darlington's 2029 target holds. Whether gas turbine delivery slots stretch past 2031. Whether the slowdown talk becomes more than a press cycle.

The first piece said the question was never whether AI needs the capacity, but who holds the risk when the music slows. The power story sharpens it: the risk isn't just financial. It's physical. You can finance a data center in a quarter. You can't build a grid connection in one, and you can't pour a reactor in two.

Money was never the constraint. The electrons were. And they're on a 2029-2033 schedule against a $700-billion-a-year construction program.

Originally published on AI & Markets: https://aiandmarkets.substack.com/p/the-ai-buildout-has-a-power-problem

The AI buildout is a balance-sheet story now

AI & Markets — Fri, 02 Oct 2026 16:22:23 +0000

The $3.1T question: who holds the risk when the music slows.

The most important number in AI right now isn't a benchmark. It's $3.1 trillion.

That's Morgan Stanley's estimate of off-balance-sheet AI commitments across nine tech firms: future data-center leases and chip purchase agreements sitting in GAAP footnotes, not on the balance sheet. About five times those companies' trailing-twelve-month capex. About three times their on-book debt and leases combined.

The AI trade stopped being a technology story a while ago. It's a financing story now: who funds the data centers, where the risk actually sits, and what breaks first.

The machine only has one direction.

The four big hyperscalers are on track for roughly $700–730 billion in combined 2026 capex, up 70–80% from last year, roughly triple 2024. Guidance keeps ratcheting up, not down. Goldman Sachs has the group approaching $1.14 trillion in 2027.

But the market is starting to ask about the return. In February, Microsoft fell 11% in a single day, its worst since March 2020, on slowing Azure growth paired with faster data-center spending. Good isn't good enough anymore; only scarce is.

The real innovation is the financing.

Because you can't spend a trillion dollars a year without getting creative about where it shows up. The structures are legal, disclosed, and invisible to every leverage metric the market watches:

Anthropic closed a $35 billion private-credit package with Apollo and Blackstone to buy TPUs through an SPV, leased back to itself. Then, on October 1, Broadcom agreed to backstop up to $42 billion of Anthropic infrastructure leases, acting as both chip supplier and lender. Anthropic's own filing flags the conflict of interest.
Oracle has ~$38 billion of similar structures tied to its $300 billion OpenAI partnership. xAI raised $20 billion through a comparable chip-collateralized vehicle.
Alphabet's contractual obligations went from $332 billion to $811 billion in a single quarter.

The pattern: the asset and the debt live in the vehicle; the tech company contributes a lease and, often, a guarantee. Legal, disclosed, in the footnotes, invisible to every leverage metric the market watches.

The test case is CoreWeave.

Strip away the structures and look at one company in the open: $35–39 billion of guided 2026 capex against $12.4–13.2 billion of revenue. Three dollars of capex for every dollar billed. Trailing free cash flow: negative $13.65 billion.

The per-contract economics genuinely work. A 100MW cluster costs ~$3.8 billion, and a five-year take-or-pay contract at $1.2 billion a year repays it twice over even if the GPUs are worth zero at the end. The problem is one level up: new five-year debt facilities that outlive the average three-year customer contract. Lenders are underwriting renewal risk that has never been tested, with the top two customers accounting for 65% of revenue.

The equity market has noticed: the stock is ~55% below its highs, Rothschild initiated with a Sell on leverage, Kerrisdale is short. The credit market, meanwhile, just handed CoreWeave an $8.5 billion investment-grade GPU-backed facility at ~5.9%. Somebody's wrong about the renewal.

The wall isn't chips. It's power.

Texas paused new data-center interconnections pending an audit of its 474-gigawatt queue, more than five times the state's all-time peak demand. The grid operator's verdict: most of it is speculative. Wood Mackenzie estimates roughly two-thirds of requested US data-center electricity will never connect. PJM's interconnection wait averages eight years.

And then there's the neighbors. More than $170 billion of data-center capacity has been blocked, withdrawn, or stalled by community opposition since January 2024. New York now publishes a framework pricing community consent at $1 million per megawatt. A Pennsylvania developer is offering residents $10,000 per household to revive a rejected project. Community pushback isn't PR anymore. It's a capex line item.

The skeptics have math.

Michael Burry's September letter put ~$3 trillion in AI infrastructure obligations on the big cloud companies and argued they depreciate AI chips over four to six years while the true replacement cycle is two to three, understating depreciation by ~$176 billion through 2028, with write-downs landing "perhaps in 2028 or 2029." His anchor stat: S&P 500 net investment hit 2.07% of GDP last quarter, the highest in roughly forty years except the March 2000 peak.

The rebuttal ran seven pages. That's how seriously the argument is being taken.

Burry's math hinges entirely on the chip-life assumption: two to three years versus the four to six the cloud companies book. The other side points to six-year-old chips still running commercial workloads. Both can't be right about the earnings. But he's not alone: Chanos sees GPU-backed debt defaults coming, Kerrisdale is short the neocloud model, and lenders are voting with their term sheets. They still won't underwrite silicon as long-duration collateral without guarantees nobody will give.

The honest version of the fiber parallel.

1998–2001 laid over $500 billion of fiber on WorldCom's myth that internet traffic was "doubling every 100 days." The real number was closer to doubling once a year. By 2002, 85–95% of the fiber was dark, bandwidth prices had fallen 90%, and more than $2 trillion of telecom value had evaporated. The glut was mocked as "100 years of capacity." It was consumed in fifteen.

What rhymes: capital flooding toward extrapolated demand, capacity overshooting, vendor financing. Lucent and Nortel financed their customers' purchases then; chip suppliers co-finance buyers now, with Broadcom acting as both supplier and lender to Anthropic.

What breaks the parallel, and it's the whole ballgame: the spenders then were junk-rated new entrants funding speculative demand with high-yield debt. The spenders now are the most cash-generative incumbents on earth, funding largely from operating cash flow against real demand. The failure mode isn't bankruptcy. It's write-downs and multiple compression.

So where does it break?

Not a prediction. A watchlist. The 2027 contract renewals, when the first 12–18-month neocloud deals reprice and renewal risk gets tested for the first time. The depreciation fight, which decides whether $176 billion of earnings were real. Whether community tolls spread beyond New York. Whether lenders ever accept silicon as long-duration collateral. Right now, they won't without guarantees the chipmakers haven't agreed to give.

The buildout is real. The demand is real. The question was never whether AI needs the capacity. It's who holds the risk when the music slows. Right now, the answer is: not the balance sheet. That's not a fraud story. It's a story about where to look.