Apple Skips the Capex Race. AI's Pricing Power Starts to Crack.

Amazon is spending roughly $200 billion on capex in 2026. Microsoft is tracking towards $190 billion. Meta raised its range to $125–145 billion. Alphabet is expecting $175–185 billion.

Apple, by comparison, is expected to spend a little over $14 billion — roughly one-tenth of even the lowest figure in that group.

It's essentially flat against fiscal 2025's spending, while the others are all ramping up capex. Apple's capital spending actually fell year over year in the December quarter, while every peer was revising upward. Combined, these four hyperscalers will likely commit about $725 billion this year, up 77% from roughly $410 billion in 2025.

At first glance, the comparison implies that Apple is sitting out on the craze, but the filings say otherwise. In the quarter ended March 28, 2026, Apple's R&D expense hit $11.42 billion, up almost 34% year over year and 10.3% of revenue, the first time Apple has crossed 10% in at least 30 years. Most of this jump is attributed to higher infrastructure and headcount costs — and Apple is spending on R&D at close to a record level. While they haven't specifically listed the costs, the timing strongly suggests AI is contributing to the increase. The difference between this spending and that of their peers is that most of Apple's isn't ending up on the balance sheet.

Apple surveyed the options, and selected Google as its principal model and cloud partner — paying a reported $1 billion a year for a custom 1.2 trillion parameter Gemini model, eight times the size of the 150 billion parameter models it had been running. Anthropic was reportedly in the running, with a $1.5 billion/year price tag.

The chosen model runs inside Apple's own Private Cloud Compute, shipped as "Siri AI" at Apple's Worldwide Developers Conference on June 8, and the agreement is multi-year and non-exclusive.

That non-exclusivity is partially the point — Apple bought this year's frontier model on a contract it can re-tender, for roughly 0.7% of Meta's total projected 2026 capex — an imperfect comparison, but one that illustrates the radically different capital intensity of the two strategies. In doing so, Apple accepts greater dependence on its largest search partner and gives up some control over the underlying model roadmap. But the upsides are impressive — it doesn't need to own such an expensive depreciating asset, and if the frontier moves meaningfully, it can more easily pivot and retender — a luxury the build-to-own players can't replicate.

This strategy isn't without roadblocks. Apple engineers reportedly tried to run Gemini on Apple's own server infrastructure and found the M2 Ultra chips — designed for high-end desktop workstations rather than hyperscale AI workloads — hit a performance ceiling on AI data center workloads. Baltra, the in-house server chip built with Broadcom, may be insufficient. Bloomberg reports a chip that can rival Nvidia may not arrive until 2029.

Apple returned $15 billion in the March quarter: $11 billion repurchasing 42 million shares, $3.8 billion in dividends. Buybacks consumed 65.2% of trailing free cash flow. Meanwhile, Meta issued $55 billion of debt in six months, Alphabet's Q1 free cash flow fell 47% to $10.12 billion, and Oracle went free-cash-flow negative and expects to raise about $40 billion in FY27 to keep building.

On May 1, Apple did something it has not done since 2018: it dropped the net cash neutral target. Net cash sits at $62 billion, $147 billion in cash and marketable securities against roughly $85 billion of debt, down from $163 billion when the policy started. Parekh's framing on the call: "Our investment in the business comes first and foremost, and then we look to return excess cash to shareholders." Evercore's Amit Daryanani read it as follows: "This is a sign they want to do more deals and invest cash differently."

Apple's largest acquisition remains Beats at $3 billion in 2014. In January it paid roughly $2 billion for Israeli AI startup Q.ai, making it the second-largest deal in company history. Per The Information, Apple has spent recent months in talks with investment bankers and semiconductor startups gauging buyouts, and is prepared to move from deals in the hundreds of millions to deals in the billions. PrismML has been named as a target for on-device processing. John Ternus, a hardware engineer, replaces Tim Cook in September, with Johny Srouji taking expanded control of all hardware engineering.

A company that has spent a decade proving it does not need to buy anything just removed its own spending guardrail, elevated its hardware people, and started calling bankers.

Chinese-origin AI models are taking a growing share of US enterprise usage, and the shift is becoming significant. In a July 7 investigation, CNBC reported that these models have accounted for more than 30% of the tokens US companies route through OpenRouter every week since February 8, peaking at 46%. The prior twelve-month average was 11%, and in the first half of 2025 it was 4.5%.

The main driver is cost. OpenRouter's data lead put the open Chinese models at 60% to 90% cheaper than the leading US systems. For high-volume, routine workloads that gap is enough to move real usage, and it has held for five straight months rather than spiking and fading.

Until now, the majority of this pressure was coming from lower-cost, less premium models — but that just changed as well. On July 16, Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model, the largest ever published. Kimi K3 scored 57 on Artificial Analysis's Intelligence Index, placing it within three points of the leading model at launch and ahead of Claude Opus 4.8. And it's not cheap — at $3 per million input tokens and $15 per million output, it is the most expensive Chinese model to date (but still relatively inexpensive by US-frontier standards). This wasn't a move to continue eroding low-cost US market share — it was intended to compete directly with the frontier.

Until now, US AI labs have primarily been competing against one another for the frontier spot, while collectively experiencing some loss of market share to Chinese open-weight models. This left them in a state of constant development, with relative freedom to price and allocate their frontier models as they pleased. K3 arrived just as the competitive pressure was becoming visible in U.S. labs' product decisions. On July 18, two days after the K3 announcement, Anthropic reversed a plan to pull its flagship Fable 5 from subscription tiers entirely, making it instead a permanent part of Max and Team Premium plans. Coverage of the reversal (The Decoder among others) tied it less to Kimi than to OpenAI's GPT-5.6 Sol, which Artificial Analysis benchmarked at roughly Fable's capability while completing equivalent tasks at about a third of the cost. Either way the direction is the same: a US lab that planned to gate its best model behind API pricing reversed its decision, ostensibly due to broader competitive implications, and widened access instead.

The scarcity premium on frontier intelligence is compressing, and that lands on deal teams two ways. If AI sits in your sourcing or diligence stack, the vendor you standardized on last year may no longer be the rational default, which makes build-vs-buy and multi-model routing live procurement questions rather than hypotheticals. And if you underwrite AI-infrastructure targets, the growth assumptions baked into those models lean on pricing power that the past two weeks called into question.