Apple Refuses $650 Billion AI Arms Race, Spends $14 Billion While Rivals Bleed Cash

Amazon, Google, Microsoft, and Meta are projected to spend a combined $650 billion on AI data center infrastructure in 2026, according to an analysis by industry analyst Horace Dediu. Apple's capital budget sits at $14 billion, nearly flat from last year. The gap between the two approaches has grown wide enough that investors and analysts are now openly debating whether Apple's restraint is discipline or denial.

Dediu described the hyperscalers' combined spending as "buying the US Navy every year." A separate Wall Street Journal op-ed by investor Daniel J. Arbess called Apple's position a "cold-eyed bet that the most frenzied build-out in the history of American capitalism will produce inadequate returns."

Cash machines become borrowers

The numbers behind that bet are brutal. Amazon, Google, Microsoft, and Meta now direct roughly 94 percent of their operating cash flows toward AI infrastructure. Amazon is projected to go negative on free cash flow this year, as much as $28 billion in the red. Alphabet's free cash flow, according to Pivotal Research estimates, is expected to collapse nearly 90 percent, from $73 billion to roughly $8 billion.

These companies used to generate more cash than they could spend. CFOs who once debated buyback sizes are now nervous about bond covenants. They borrow to keep building. The Big Five raised $121 billion in bonds in 2025 alone. Morgan Stanley projects $1.5 trillion in tech debt over the coming years. Hyperscalers, for the first time in their history, hold more debt than cash.

And the returns? AI services generate roughly $35 billion in total revenue across all providers. Five percent of what the industry is spending on infrastructure.

Apple spent $90.7 billion buying back its own stock last fiscal year. Its competitors' combined buybacks collapsed 74 percent from their peak as cash got rerouted to data centers.

Why renting a model beats building one

Rather than building its own foundation models, Apple licensed Google's Gemini for a reported $1 billion annually, about 1 percent of its free cash flow. The Trefis team at Forbes put it plainly. Why amortize a $100 billion infrastructure build when outsourcing costs a fraction? If something better surfaces from Anthropic, DeepSeek, or the open-source community, Apple switches vendors.

That flexibility matters more as models commoditize faster than most predicted. Dediu pointed to DeepSeek, which built a model for $6 million that matches systems costing $100 million. Open-source models, by his estimate, now power 80 percent of startups seeking venture funding.

Apple is not sitting idle on silicon, though. The M5 chip, announced last October, embeds neural accelerators into every GPU core. Arbess noted that Apple's own benchmarks show it running a 30-billion-parameter model in under three seconds on a standard MacBook Pro. No internet. No subscription. The architecture fires only three billion parameters per query out of the full thirty billion, a Mixture of Experts design that keeps inference fast and local.

2.5 billion devices as the data center

Apple has 2.5 billion active devices in the field. That number changes the math. Every iPhone, Mac, and iPad becomes a distributed inference node, handling queries that would otherwise hit a server farm. If you ask Siri to summarize an email or edit a photo, that task runs on silicon in your hand. It never touches a data center.

Arbess singled out Meta as the most exposed company in this reordering. Meta has no platform layer, no operating system, no cloud business, no device. It spends up to $135 billion this year to build consumer AI features that Apple ships free at the operating system level. A query Siri resolves on-device is a session that never begins on Instagram. And when an AI notification summary compresses twenty minutes of scrolling into thirty seconds, that is twenty minutes of ad impressions Meta never serves. The company looks emboldened by its ad revenue growth, but Apple is quietly draining the attention underneath it.

The risk nobody dismisses

If a category-defining AI application emerges that demands massive server-side compute, Apple's conservative infrastructure spend leaves it cornered. Enterprise AI workloads, multiagent systems orchestrating millions of documents, require data-center hardware no laptop can match. Nvidia's Vera Rubin platform delivers 36 times the memory bandwidth of Apple's best chip.

Apple's dependence on Google for backend AI carries its own exposure. Regulatory scrutiny of the partnership remains a live risk. And Apple Intelligence's track record since launch has been, charitably, uneven. Daring Fireball's John Gruber captured the tension: making Apple Intelligence a first-class agentic AI by relying on Gemini at $1 billion a year "sure looks like genius. But given their track record with Apple Intelligence to date, that is an enormous 'if.'"

Investors pick the company that kept its cash

Wall Street has already voted with its positioning. Apple's stock is positive year-to-date. Microsoft and Amazon have dropped 18 and 13 percent respectively since January, according to a MarketMinute analysis. Institutional investors rotated out of speculative AI infrastructure plays and into Apple as a defensive hold. Apple's Q1 fiscal 2026 results, $143.8 billion in revenue and a 16 percent jump, reinforced the thesis.

Either AI models continue to commoditize and shrink onto local hardware, vindicating Apple's patience. Or centralized compute becomes the bottleneck for the next generation of applications, and Apple finds itself renting capacity from the very companies it watched spend their way into debt.

The $90.7 billion in buybacks says Apple is not hedging.

Marcus Schuler

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]