AI economics are tilting hard toward GPUs and fiber: labs and hyperscalers are pre-buying massive compute even as cheap local models get good enough for most everyday work. At the same time, Apple and Google are fusing the mobile AI stack while OpenAI and Anthropic race to IPO into a market that’s already showing bubble fatigue.
The real spread is between players pricing in those future constraints now and those still behaving like it’s a limitless land grab.
Key Events
/Apple unveiled Siri AI on a new architecture using Google Gemini models, with rollout blocked in the EU and China under the DMA.
/Anthropic agreed to pay $1.25B for 220,000 GPUs at Colossus 1, one of the largest single AI compute purchases disclosed.
/Google struck a cloud deal to pay SpaceX about $920M per month for computing services.
/OpenAI confidentially filed for a U.S. IPO while facing a first-of-its-kind safety lawsuit from Florida over ChatGPT.
/Massachusetts passed a privacy rights law banning the sale of precise location data.
Report
Two things moved this month: the AI cost stack flipped from people to compute, and mobile AI power consolidated around an Apple–Google axis. Everything else—IPO hype, data center politics—is downstream of those shifts.
the new ai cost stack: compute > payroll
Nvidia’s VP says AI compute now costs more than the engineers, with Uber burning through its entire annual AI budget by April on model tokens.
Uber engineers report monthly AI token charges in the low thousands of dollars, leading leadership to drop token usage as a performance KPI.
The five largest U.S. tech firms spent about $380B on capital expenditures in 2025, largely driven by AI infrastructure.Anthropic committed $1.25B to GPU purchases at Colossus 1, while Google is paying $920M for a large GPU tranche, effectively fixing their compute costs well ahead of demand.
A Wharton analysis says AI must raise productivity by 2.7× to avoid mass tech bankruptcies, just as a Stanford study finds local models now solve most everyday queries that previously required frontier cloud models.
mobile ai realignment: apple + google as one stack
Apple announced a new AI architecture built around Google Gemini models, with its next‑gen Apple Intelligence and Siri AI effectively running on distilled Gemini.
Apple says its AFM Core Advanced model has 20 billion parameters yet can run on iPhones via the A19 Pro chip, something that usually requires much larger hardware.
Siri AI’s more conversational, context‑aware features are rolling out via a waitlist in English, but are blocked in the EU and China under the Digital Markets Act, delaying iOS 27 and iPadOS 27 upgrades there.
The DMA also forces Apple to allow equal access for third‑party AI providers, turning Siri into a contested distribution surface and complicating Apple’s preference for tightly controlled, on‑device intelligence.
Commenters are openly skeptical that Apple is innovating here, accusing it of rebranding Google tech, questioning whether privacy promises survive a Gemini dependency, and worrying that OpenAI and Anthropic are being boxed out of the default mobile surface.
ai infra hits physical and social limits
A Texas town sold 87 acres of land originally donated for a park to a data center developer for $10M, crystallizing how AI infrastructure is now outbidding civic uses.
Working‑class neighborhoods are resisting data centers at roughly five times the rate of wealthy areas, even as many new AI facilities are sited on drought‑hit land that is already short of water.
Texas grid officials have warned that data centers and crypto sites are failing voltage tests and creating system risks, turning power reliability into a gating factor for AI expansion.
At the same time, Amazon engineers in Seattle are protesting billion‑dollar AI data centers and a fiber deal with Corning while 30,000 staff are laid off, and Meta is throwing up tent‑based chip farms to chase scarce compute.
With shipowners exploring floating data centers and the UK government planning a £1B spend to boost public compute capacity twentyfold, the scramble is shifting from land availability to water, grid headroom, and long‑haul fiber.
capital markets, frontier labs, and sovereign demand
OpenAI has confidentially filed for an IPO just as Florida becomes the first U.S. state to sue Sam Altman and the company over ChatGPT’s societal risks, putting safety and concentration of power onto the S‑1.
Reports put Anthropic’s market value near $965B, prompting questions about how much of that is backed by real cash flows.
Commenters cite figures around $47B in annual revenue for Anthropic and broader worries that upcoming AI IPOs are overvalued and their business models unproven.
The broader AI sector has already seen about $1.3T in market value erased in a wave of profit‑taking and bubble fears, with SpaceX’s planned IPO labeled overvalued by Morningstar and short sellers like Steve Eisman openly skeptical.
Against that, governments and defense actors are becoming anchor customers—Anthropic’s new Mythos model is reportedly used by the U.S. Department of War for cybersecurity, and UK and U.S. public funding is flowing into frontier compute and self‑improving AI labs.
What This Means
The live decision is whether to play the AI land‑grab—locking in expensive compute, riding Apple–Google’s mobile stack, and leaning into frothy IPOs—or to assume a correction and design for a world where cheaper local models, physical constraints, and security failures cap returns.
Whichever way it breaks, the spread between those who priced in real AI economics early and those who didn’t will be where most of the money moves.
On Watch
/Predictions from within the AI community of a major agent-related disaster within a year, as access to sensitive systems outpaces guardrails.
/Shipowners and Samsung Heavy exploring floating data centers as land, water, and grid constraints bite onshore capacity.
/Stanford’s finding that local models now correctly answer 71.3% of real‑world queries, up from 23.2% in 2023, which could accelerate a shift away from expensive frontier APIs.
Interesting
/The five largest U.S. tech firms are projected to double their capital expenditures from approximately $380 billion in 2025 to an estimated $760 billion in 2026.
/Moonshot AI, a Chinese lab, is seeking $30 billion in funding amid fierce competition in the AI sector.
/AI voice agents are rapidly replacing traditional call centers, driven by advancements in conversational AI and workflow automation.
/HP announced the most powerful Windows AI PC ever built, capable of handling one trillion parameters with 784GB of unified memory.
/Jensen Huang's visit to Seoul was focused on securing Nvidia's role in Korea's industrial stack rather than chip sales.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/Apple unveiled Siri AI on a new architecture using Google Gemini models, with rollout blocked in the EU and China under the DMA.
/Anthropic agreed to pay $1.25B for 220,000 GPUs at Colossus 1, one of the largest single AI compute purchases disclosed.
/Google struck a cloud deal to pay SpaceX about $920M per month for computing services.
/OpenAI confidentially filed for a U.S. IPO while facing a first-of-its-kind safety lawsuit from Florida over ChatGPT.
/Massachusetts passed a privacy rights law banning the sale of precise location data.
On Watch
/Predictions from within the AI community of a major agent-related disaster within a year, as access to sensitive systems outpaces guardrails.
/Shipowners and Samsung Heavy exploring floating data centers as land, water, and grid constraints bite onshore capacity.
/Stanford’s finding that local models now correctly answer 71.3% of real‑world queries, up from 23.2% in 2023, which could accelerate a shift away from expensive frontier APIs.
Interesting
/The five largest U.S. tech firms are projected to double their capital expenditures from approximately $380 billion in 2025 to an estimated $760 billion in 2026.
/Moonshot AI, a Chinese lab, is seeking $30 billion in funding amid fierce competition in the AI sector.
/AI voice agents are rapidly replacing traditional call centers, driven by advancements in conversational AI and workflow automation.
/HP announced the most powerful Windows AI PC ever built, capable of handling one trillion parameters with 784GB of unified memory.
/Jensen Huang's visit to Seoul was focused on securing Nvidia's role in Korea's industrial stack rather than chip sales.