US$811 Billion & Falling: Alphabet is Burning Its Way to Irrelevance

The US AI arms race is not a strategy.  It is a spending competition funded by debt, consuming electricity at the rate of a small country, and competing against an adversary that has already proven it can do more with less.  Alphabet’s Q3 2026 results are not an anomaly.  They are a preview.

Alphabet posted Q3 2026 revenue of US$127.6 billion.  Analysts had expected US$128.2 billion.  The shortfall was $600 million — less than one per cent of revenue, and in a normal quarter, forgettable.  This was not a normal quarter.  Adjusted earnings per share were US$2.67, against a Wall Street target of US$3.02.  That is a 12 per cent miss.  Net income fell 25 per cent year on year.  Free cash flow was negative US$9.07 billion — worse than the negative US$5.9 billion posted in Q2 2026, which was itself the first negative quarterly cash reading in at least a decade.  The cash burn is accelerating.  Shares fell more than 10 per cent in extended trading.

Pichai Sundararajan, Chief Executive Officer of Alphabet, told investors that Google was “leading at the frontier.”  He said this after the worst earnings report the company has produced in years.  He was awarded a pay package worth up to US$692 million over three years in March 2026.  US$84 million of that vests simply for remaining in his seat.  One presumes the board finds the current trajectory acceptable.

The Infrastructure Commitment: US$811 Billion in Obligations

The annual capital expenditure number is alarming enough.  Alphabet raised its 2026 capital expenditure forecast to US$195–US$205 billion, up from US$180–US$190 billion set just months earlier.  Chief Financial Officer Anat Ashkenazi confirmed that AI-driven capital expenditure will rise further in 2027.  She gave no figure.  The more alarming number is the forward spending commitment.  At the end of June 2026, Alphabet’s contracted future spending commitments reached US$811 billion — up nearly US$500 billion in three months.  These are not ambitions.  They are legal obligations under supply agreements and open purchase orders: chips, data centres, electricity contracts, content licences.  Alphabet is committed to spending US$811 billion.  For perspective: Singapore’s entire gross domestic product in 2025 was US$560 billion.  Alphabet has contractually obligated more than that to its AI infrastructure programme.

The five hyperscalers together — Alphabet, Amazon, Meta, Microsoft, and Oracle — have committed about US$969 billion, per Moody’s analysis.  US$662 billion of that consists of data centre leases not yet commenced.  Moody’s notes that under standard accounting rules, leases only appear on the balance sheet when renewal is “reasonably certain,” which means most of these obligations are invisible to a casual reading of the accounts.  Goldman Sachs projects US$1.15 trillion in hyperscaler capital expenditure for 2025–2027 combined.  The Big Five alone are projected to spend US$602 billion in 2026.  That is a 36 per cent increase from 2025.

The Debt Cliff

Until recently, Alphabet had negligible long-term debt.  At the end of 2024, it carried US$10.9 billion.  By the end of 2025, that had risen to US$46.5 billion — a 327 per cent increase in twelve months.  In February 2026, Alphabet sold another US$20 billion in bonds, including, with breathtaking confidence, a 100-year sterling bond.  Total debt now stands at US$70 billion.

Alphabet is not alone.  Technology companies issued a record US$428 billion in bonds in 2025, up 95 per cent from 2024.  Hyperscalers added US$121 billion in new debt in 2025, with over US$90 billion issued in three months.  Morgan Stanley projects AI-linked bond issuance of nearly US$570 billion in 2026.  By October 2025, AI-linked debt had overtaken US banks as the largest investment-grade sector by issuance volume.  Outstanding AI-related debt tracked by M&G Investments stood at US$1.2 trillion.

The interest arithmetic is sobering.  At US$121 billion in new debt and prevailing rates of 5–6 per cent, the incremental annual interest burden reaches US$6–US$7 billion.  IBM Chief Executive Officer Arvind Krishna has calculated that US$8 trillion in total AI capital expenditure would require US$800 billion in annual profit just to cover interest — not to return capital, just to service the debt.  Nobody has explained where that profit comes from.

The outlier in the group is Oracle.  Barclays’ fixed income analyst Andrew Keches, in a November 2025 report, downgraded Oracle’s debt to Underweight and warned it could eventually fall to BBB- — near junk.  Oracle carries a debt-to-equity ratio of 500 per cent, against Amazon’s 50 per cent and Microsoft’s 30 per cent.  It is the only hyperscaler already running negative free cash flow on a structural basis.  Off-balance-sheet lease commitments exceed US$100 billion.  Barclays’ stress test concluded that Oracle’s cash could run out by November 2026.  Its US$300 billion contract with OpenAI has made its credit default swaps a proxy for OpenAI risk, and Oracle’s five-year CDS spreads have more than tripled since September 2025.

S&P has warned that Amazon will likely post negative free operating cash flow over the next two years.  PIMCO’s Chief Investment Officer for Core Strategies, Mohit Mittal, has warned of possible overinvestment over the next two years and a correction.  Passive funds managing US$4.8 trillion in target-date retirement assets must absorb eligible AI bonds as they are issued.  A credit event in the sector would ripple into millions of pension accounts.  This is not a technology story.  It is a credit story.

The Revenue Problem: Spending to Destroy Your Own Business

Alphabet earns most of its revenue from advertising.  Core advertising revenue in Q3 2026 was US$82.01 billion — US$1.41 billion below expectations.  The advertising business has historically enjoyed near-monopoly economics: Google held more than 90 per cent of global search market share for most of the past decade.  In December 2024, that share fell below 90 per cent for the first time in nearly a decade — to 89.73 per cent globally and 87.39 per cent in the United States.  ChatGPT, Perplexity, and xAI are the beneficiaries.  People are asking AI tools for answers instead of searching for links.  The ten-link results page, the format that generates Alphabet’s advertising revenue, is losing relevance.

This is the central irony.  Alphabet is deploying AI-powered search that cannibalises its own advertising revenue.  It is then borrowing US$70 billion to build the infrastructure for an AI revenue model that does not yet exist at scale.  Alphabet’s own annual filings warn that AI-powered search could threaten its 90 per cent-margin advertising business.  It is spending US$205 billion to accelerate a transition it cannot yet monetise, funded by debt it cannot yet service from AI revenue.

Google Cloud — the business that rents out AI computing power — is the exception.  Cloud revenue more than doubled to US$30.71 billion in Q3 2026.  Cloud grew 82 per cent.  It is also the business that requires Alphabet to spend US$205 billion on infrastructure to fulfil demand, at margins that are a fraction of the 90 per cent advertising margin.  Alphabet is trading a near-monopoly, high-margin business for a capital-intensive, competitive, lower-margin one.  Pichai described this as “leading at the frontier.”  The frontier is burning US$9 billion a quarter in free cash flow.

The historical parallel is Eastman Kodak Company.  Kodak invented the digital camera in 1975 and shelved the technology to protect film revenue.  Competitors developed digital photography.  Kodak filed for bankruptcy in January 2012.  Kodak was protecting a profitable business at the cost of missing the future.  Alphabet is destroying a profitable business and spending US$205 billion on a replacement that has not yet been built.

The Strategic Landscape: America’s Expensive Mistake

The hyperscaler strategy rests on a premise: build the largest clusters, purchase the most chips, consume the most electricity, and win by volume.  The Big Five are projected to spend over US$700 billion collectively in 2026 on this premise.  That premise was invalidated in January 2025.  DeepSeek, a Chinese artificial intelligence firm, released a model trained for under $6 million using 10,000 Nvidia chips.  It scored 92 per cent accuracy on logical reasoning benchmarks, against GPT-4’s 78 per cent.  It was released as open source.  Nvidia Corporation lost US$593 billion in market capitalisation in a single trading session — the largest single-day loss of market value in stock market history.  OpenAI spends over US$5 billion annually.  DeepSeek spent US$6 million.  The US response was to raise capital expenditure targets.  This is the strategic equivalent of learning that a competitor crossed the Atlantic in a rowing boat, then commissioning a larger ocean liner.

The mechanism behind DeepSeek’s efficiency was the US export controls on advanced chips.  By restricting Nvidia’s H100 and H200 to Chinese buyers, the US forced Chinese engineers to work within hardware constraints.  Constrained hardware demands algorithmic efficiency.  Algorithmic efficiency produces models that run on less compute.  The policy achieved precisely the opposite of its stated objective.  Huawei’s Ascend 910C, the domestic Chinese AI accelerator developed in response to the export controls, delivers 60 per cent of the inference performance of Nvidia’s H100, per DeepSeek’s own benchmarks.

Domestic Chinese chips now account for 41 per cent of China’s AI market, up from single digits before 2023.  China’s 15^th Five-Year Plan mentions AI over 50 times and commits to “decisive breakthroughs in chips.”  China prepares to invest US$295 billion in nationwide AI data centre infrastructure over five years, through state operators China Mobile and China Telecom, using at least 80 per cent domestic chips.  Alibaba’s open-weight Qwen model has overtaken Meta’s Llama in global download volumes.  Silicon Valley startups build on Chinese open models.  China added over 540 gigawatts of new power capacity in 2025 — more than enough to support rapid data centre construction without the grid stress afflicting the United States.

The Brookings Institution, in its analysis of competing AI strategies, concluded that the winner will be determined “not simply by who builds the most powerful models” but by who best deploys AI across the economy.  China is deploying AI into manufacturing, logistics, agriculture, and government services by state directive — no subscription revenue required, no market signal needed.  The United States has better foundational models.  China has a larger installed base of industrial applications and a fraction of the debt load.

The Energy Constraint

US data centres now consume 7 per cent of national electricity demand, per the International Energy Agency — up from 4.4 per cent in 2023.  Data centres drove half of US electricity demand growth in 2025.  The IEA projects that pace to continue through 2030.  The grid is not ready.

The North American Electric Reliability Corporation warns of elevated risk of summer electricity shortfalls across all three major US grid regions through 2026 and beyond.  Utilities in Virginia, Texas, Oregon, and Arizona face generation and transmission backlogs that are delaying facilities already financed and permitted.  The PJM electricity market — covering Illinois to North Carolina — recorded an estimated US$9.3 billion rise in capacity market costs for 2025–26, adding US$16–US$18 per month to residential electricity bills in parts of Ohio and Maryland.  At least six states introduced bills to pause new data centre construction.  Residents near major hubs in Virginia, Texas, and Georgia are opposing new facilities.  Northern Virginia developers face permit queues until 2032.  A single large data centre facility consumes as much water as 6,500 households.  The US strategy requires limitless cheap power and limitless public tolerance.  Both assumptions are failing simultaneously.  China added 540 gigawatts of new capacity in 2025.  The US added 50 gigawatts.  The infrastructure competition is not close.

The National Interest Calculation

The strategic argument for the hyperscaler spend is that AI dominance is a national security matter.  That argument has merit.  The execution is a different question entirely.  Spending US$811 billion in committed obligations on infrastructure for a technology whose revenue model is unproven is not a strategy.  It is a bet.  The distinction matters when the bet is financed by US$1.2 trillion in outstanding AI-related debt, when the underlying revenue business faces structural erosion, and when the adversary has demonstrated it can match the output at 0.1 per cent of the cost.  Hyperscalers currently spend close to 100 per cent of operating cash flow on capital expenditure, against a ten-year average of 40 per cent.  That is not investment.  That is desperation.  The market priced that desperation at negative 10 per cent on Alphabet’s Q3 2026 results.  It will price it again.

The race is not between GPT-5 and whatever Alibaba releases next.  The race is between a model that costs US$5 billion a year to operate and one that costs US$6 million to build.  The mathematics of that competition do not favour the side carrying US$1.2 trillion in debt.  Pichai Sundararajan is “leading at the frontier.”  The frontier, at the current rate of cash burn, has a finite runway.

---

Terence Nunis | Executive Chairman, Equinox Zenith | Author, The 1% Playbook: The Billionaire Cheat Code