Chips That Cannot Be Plugged In: Inside the $1.7 Trillion AI Bill

What if the largest misallocation of capital of our lifetime is already sitting in warehouses, paid for but never switched on? That question frames a Market Talk debate where host George Noble brings together Ed Zitron, Gary Marcus, Julien Garran, Jack Gamble and Alessandro for ninety minutes on whether the AI boom is history's priciest resource sink — and who foots the bill if it breaks.

Gary Marcus opens by hitting the brakes on the rogue AI panic. There is no outlaw intelligence on the loose, he argues, only badly engineered software. His analogy lands hard: if you build a car with cardboard brakes and it runs out of control, you do not call it a rogue car, you call it bad engineering. Claims of autonomous agents going off-script are, in his view, the same fallacy — a system that cannot follow instructions is behaving exactly as designed.

The second act of the fear narrative features air-gapped computers and models escaping the lab. Marcus says stories of frontier systems outsmarting technicians trying to shut them down are overblown. Large language models cannot reliably follow everyday instructions; ask for an email summary and get something else entirely. Against that reality, tales of models hatching secret plans look less like technical reporting and more like a mythology brewed from marketing and panic.

Broken models or broken engineering?

The alleged Navier-Stokes proof solved by OpenAI's Codex coding assistant sparks one of the panel's fiercest exchanges. The story goes that a researcher's open-source work was presented as if a chatbot had single-handedly cracked a mathematical problem beyond human reach. Gamble calls the framing an appropriation of intellectual labor: taking years of someone else's work and putting your own product in the shop window. For him, the episode distills how the industry manufactures scientific prestige.

Ed Zitron aims his fire at the numbers. Until independently audited, he says, neither Anthropic's nor OpenAI's revenue and growth figures deserve trust. The annualized revenue run-rates they parade are cooked differently at every company; some multiply 28 days by thirteen, others a single month by twelve. That flexibility polishes investor decks while hiding real cash flow, Zitron argues, and only audited financials can cut through it.

The leaked Anthropic IPO filing feeds those doubts: documents show roughly 8 billion dollars in operating losses next to dizzying revenue growth. A debated 2 trillion dollar valuation sits awkwardly beside that red ink, especially amid talk that the company caught up with rivals only after changing its calculation method. This picture is laid out line by line in the leak analysis published on siliconangle.com, and read together with the IPO reporting from reuters.com the story comes full circle.

The least discussed layer of the AI stack is Broadcom and Google's custom silicon . The panel stresses that risk is piling up not only on Nvidia's doorstep but in the specialized accelerators entering data centers and the networking gear stitching them together. Google's TPU push may look like reduced dependence on a single supplier, yet it does not remove systemic fragility; instead, trillions in spending become the same wager spread across different balance sheets.

1.7 trillion dollars already spent

The toughest arithmetic of the night comes from MacroStrategy partner Julien Garran. Big Tech has already sunk about 1.7 trillion dollars into AI capital expenditure, with plans calling for another trillion a year. Covering the depreciation burden of past investments plus interest costs would require roughly a trillion dollars in additional annual earnings — meaning these companies must essentially reinvent their entire businesses. That projection lines up with the research note published on goldmansachs.com forecasting global AI investment to top 1 trillion dollars in 2026.

Even on a generous depreciation schedule the picture is frightening: by Garran's math, AI investments wipe out 98 percent of Big Tech's non-AI income by 2034. The critical distinction is revenue versus profit; 1.3 trillion dollars in compute commitments is not money in the till but lease-like obligations falling due in the future. With Microsoft's operating expenses alone running near 150 billion dollars, how those commitments meet actual cash flow remains unanswered.

The dot-com dark fiber analogy fails this time, the panel argues. Fiber laid in that era could wait; it held its value and lit up when demand arrived. Graphics processors , by contrast, lose value every year, consume electricity, and become unrentable two generations later. A chip waiting unused is not waiting cable but melting ice, so anyone drawing lessons from the last crash should read the analogy in reverse. This depreciation debate is examined from every angle in the dossier published on cnbc.com on the economic lifespan of AI graphics chips.

Meta's Muse assistant walks on stage as living evidence of why the killer app still has not shown up. On paper, an autonomous agent that places phone calls sounds impressive; in practice, 404 Media reported that some calls were quietly handled by humans in a call center. For Zitron, the episode exposes the gap between generative AI demos and field reality: the shop window is autonomous, the kitchen is manual. That reporting, documenting human involvement during testing, was published on 404media.co.

Agents, banks and the triggered stampede

The scenario of autonomous agents plugged into the financial system is the panel's darkest fantasy. If communicating software agents all rush to close positions at once, a crypto-flavored algorithmic bank run could follow. Cascading sell orders drain liquidity while margin calls ripple outward wave after wave. The participants insist this is no distant dystopia; in leveraged, interconnected markets, a single trigger is enough to start a machine-speed stampede.

The final address on the bill, though, is the ordinary saver. The panel says risk has been parked in life insurance policies, pension funds and 401(k) accounts, carried unknowingly by the viewer. As private credit floods data center projects, an estimated 800 billion dollars in venture capital that flowed into AI startups since 2022 is expected to go largely to zero. So if the bust comes, it will show up not on the stock ticker but on the retirement statement.

Amazon's plan to sell about 8 billion dollars of Nvidia chips and lease them back is dissected as the showcase of creative financing. Moving Grace Blackwell chips into a special financing vehicle and renting them back pushes debt off the balance sheet without removing the obligation. Gamble reads it as balance-sheet engineering that veils the gap between real profitability and presented health. The structure is laid bare in the analysis published on techrepublic.com costing out the sale-leaseback model.

Rents rising while chips gather dust

Rising chip rental prices cast a cross-light on the picture. According to market trackers, Nebius H100 rental rates climbed markedly at the start of October 2026; tight supply pushing prices up seems at first glance to contradict claims of idle hardware. The panel resolves the paradox: rented capacity and purchased-but-unpowered capacity are not the same thing. Those price moves are plainly visible on the dashboards published at gpueconomy.com that track hourly rental rates day by day.

The night's headline belongs to Microsoft: the company is said to hold some 80 billion dollars in AI chips it cannot plug in, queued behind grid connections and power supply. Data center construction may be outrunning chip purchases, but a rack without electricity is just capacity on paper. Zitron's research scales the figure up further, estimating 200 to 350 billion dollars in graphics processors sitting idle in warehouses and unpowered facilities. That power bottleneck is examined in depth in the dossier published on datacentremagazine.com explaining why chips are stuck in inventory.

The idle-hardware claim meets the merciless math of fixed costs. In a landscape where Chinese rivals produce cheaper and fat-margined giants pass costs to customers, price pressure looks inevitable. Marcus recalls that graphics processors touted to last a decade exhaust their economic life far sooner in practice; an aging chip can neither be sold nor rented competitively. Each new generation thus works like a tax that scraps its predecessor.

The prisoner's dilemma and the gold question

Big Tech's position is a textbook prisoner's dilemma: everyone keeps spending for fear that whoever stops first loses, though a collective pause would serve all. Noble cuts in with a lesson from his shipping years, when everyone ordered vessels, freight rates collapsed, and everyone paid the bill anyway. The capital cycle never changes, he says, only the ships have become data centers; in the end oversupply grinds down prices and paper profits with equal harshness.

On monetary policy, the question is what a bust does to gold, the dollar and Treasuries. Gold stands out as the traditional refuge in such confidence crises, while long-dated Treasury bonds will decide the fate of the yield curve. The Fed's room for maneuver is squeezed between inflation and financial stability, and the dollar's reserve privilege gets tested not at the first tremor but in the absence of alternatives. The panel differs on which inning this is, yet agrees on the wind's direction: it is turning against them.

Toward the close, Marcus points at the market: AI stocks now trade on flows rather than fundamentals, a setup that reminds him of the meme-stock era. Nobody can date the break, but the mechanism gets spelled out: private-market valuations seize up first, public equities wobble next, and the credit tap tightens last. The sequencing resembled dot-com; the difference is that leverage now flows through more hidden channels.

Zitron's closing draws a loop: inflated valuations attract fresh investment, fresh investment feeds inflated expectations, and nobody hits the brakes until an audit arrives. Only verified figures and cash-tested commitments can break this doom loop , he argues. Noble gets the last word: viewers should leave not with fear but with a question — where their money is parked and whose word it rests on. Because when the bill comes due, it will not be the faces on screen who pay, but the faces watching them.