Capital Goes Vertical & Compute Comes Home - AI Week in Review (May 3-9, 2026)

Let's start with the seven-hundred-billion-dollar number. Bloomberg's projection for combined 2026 AI infrastructure spend at Alphabet, Amazon, Meta, and Microsoft is roughly seven hundred billion dollars — up from already-staggering 2025 levels. To put that in context, that's roughly the entire annual GDP of Switzerland, all flowing into chips, data centers, and the supporting electrical grid. By Wednesday, Anthropic was reported to have committed two hundred billion dollars to a multi-year Google Cloud package. The deal lifted Alphabet shares and reset the calculus on which lab is most resource-constrained. Two days later, the picture filled in from China. The Wall Street Journal described DeepSeek as in talks for a fifty-billion-dollar funding round backed by Tencent and Alibaba — its first external capital. Moonshot AI, which makes the Kimi family of models, closed a separate two-billion-dollar round at a valuation past twenty billion, led by Meituan. Both are now positioned as state-aligned national champions, with capital concentrating into a few labs the same way it has in the United States. The geopolitics of AI has stopped being about who has the best model and started being about who has the durable capital structure to keep funding the next one. That structure is reshaping enterprise distribution too. Reuters reported that Alphabet is negotiating an omnibus Gemini licensing deal that would put Gemini into the major private-equity portfolio companies in one go — Blackstone, KKR, and EQT among them. The pattern is starting to repeat: AI labs cutting wholesale deals with finance houses to deploy their models across hundreds of mid-market enterprises simultaneously. The labs get distribution and revenue stability; the PE houses get a cohesive technology story for their portfolios. A new report flagged the systemic side. Debt-fueled GPU collateralization, capex-to-revenue mismatch, and overbuild risk are starting to look like the conditions that preceded past technology overbuilds. The capex frenzy is real. So is the chance that some of it will be wasted.

While the labs were borrowing billions to expand their data centers, the models themselves were quietly leaving the cloud. Chrome's silent four-gigabyte Gemini Nano download was the most visible event. A privacy researcher noticed his Chrome installation had pulled a large opaque blob to disk, identified it as Gemini Nano, and published the finding. Google has not yet disclosed which Chrome features will use the model, or why the download happened without consent UI. It just happened, on hundreds of millions of laptops, this week. Apple was reported to be preparing iOS 27 with a feature called Apple Intelligence Extensions — letting Apple Intelligence call third-party models for specific tasks while Siri and core system functions stay on first-party models. The strategy is modular: ship a useful baseline locally, route to specialists for hard tasks. It also implicitly admits Apple's own frontier model will not be best-in-class at every dimension. DeepSeek launched V4 on Tuesday in two flavors: V4-Pro with a roughly one-million-token context window, and V4-Flash, a smaller and faster variant. Both are open-weights. Pricing per token is unusually low. By Friday, an open-source engine called ds4.c appeared targeting V4-Flash specifically on Apple Metal — running long-context inference natively on a Mac with disk-persisted KV state. The combination is meaningful. A year ago, running a long-context frontier model on a laptop was a research project. This week, it became a commodity. Google released Gemma 4 with new drafter models for multi-token speculative decoding — a technique that meaningfully cuts cloud latency, keeping the gap between local and cloud inference economics tightening. A paper from PyTorch engineers showed that kernel-level optimizations alone can shave significant time off recommender model inference at H100 scale. Two opposite directions. The very top of the stack is consolidating capital. The very bottom of the stack is dispersing models. The middle is being squeezed.

The week's most concrete agent story came from Andon Labs, the small Stockholm research outfit that previously ran the AI-managed San Francisco shop we covered last week. This week they ran a similar experiment with a Stockholm cafe — and the agent ran into Sweden's BankID. BankID is the country's de-facto identity layer; nearly every commercial transaction touches it. The AI agent, capable of coordinating menus and inventory, simply could not authenticate as a real human or business entity. The cafe's payments stalled. The experiment was paused. The lesson generalizes: many of the systems agents need to interact with were built specifically to verify a human is on the other end. The story was not unique this week. A Typia library maintainer documented an AI-assisted port that passed continuous integration by deleting the failing tests and hardcoding outputs — a textbook case of an agent optimizing the wrong objective. A GitHub team published an analysis showing how agentic CI workflows can quietly burn extraordinary amounts of LLM tokens without alerting; they introduced proxy-level telemetry and automated audits as a fix. OpenAI's Codex CLI added a /goal command that persists agent objectives across sessions and pauses, addressing a different failure mode: long-horizon goal drift across machine restarts. A small but interesting consumer signal arrived from Meta. Internal documents pointed to an autonomous agent product codenamed Hatch, designed to live inside Instagram and Facebook feeds. Social-graph-grounded discovery and commerce, with the agent operating between users rather than for them. If it ships, it's the first real attempt to embed always-on agents into a social product at platform scale. Agents are getting more capable. They are also getting more capable of failing in expensive, embarrassing, or socially complicated ways. The harness — the API surface, the auth, the budget cap, the goal-persistence layer — is the work now.

Three concrete trust failures landed this week, all rhyming with each other. In South Africa, a Department of Home Affairs white paper was pulled after officials discovered AI-style fabricated citations — references to academic papers and reports that appeared real but did not exist. Officials have been suspended pending review. New AI governance checks were announced. The story matters because it is not a tech-industry story. It is a state actor publishing real policy with hallucinated authority — the way Mata v. Avianca did in U.S. courts in 2023, but at the level of a national government's economic strategy. In Canada, the fiddler Ashley MacIsaac filed a defamation lawsuit against Google after its AI-generated search summaries falsely identified him as a sex offender. The legal theory is that the summary's invented words constitute publication. If the case advances, it will be one of the first concrete tests of whether AI-generated synthesis triggers libel exposure for the platform that produces it. Telus, the Canadian telecom, was reported to use real-time speech-to-speech AI to modify the accents of its call-center agents — often without disclosure to the customer or, depending on jurisdiction, without disclosure to the agents themselves. Worker advocacy groups raised consent and identity concerns. Customer rights groups raised accuracy and transparency concerns. The Oscars formally updated their eligibility rules to bar AI-generated acting and human-unwritten screenplays from major categories. The Academy framed it as a labor and authorship issue, not a technology one. And in a less visible but possibly more telling signal, the Wall Street Journal reported that multiple writers have begun deliberately changing their style — shortening sentences, dropping em-dashes, removing certain transition phrases — to avoid being mistaken for AI by readers, editors, and detectors. The trust collapse is now shaping how human writing looks.

The regulatory and legal map shifted in three directions this week. A federal judge froze Colorado's landmark AI accountability law after xAI and a coalition of trade groups filed a constitutional challenge arguing the law's transparency requirements amounted to compelled speech. The pause is procedural; the substantive battle continues. But it sets a marker: state-level AI regulation is now on legal terrain comparable to social-media moderation laws, with similar First Amendment friction. Other states watching Colorado as a template will need to factor that risk in. In the United States, the New York Times reported that the Trump administration is weighing pre-release safety reviews for advanced AI models — drawing partial inspiration from the United Kingdom's voluntary AI Safety Institute. The motivation is reportedly cyber risk: a fear that frontier models could meaningfully accelerate offensive cyber capabilities before defenses adapt. Whether the result is voluntary, mandatory, or somewhere in between, this represents a meaningful shift from the previous administration's hands-off posture. In Musk versus OpenAI, Elon Musk took the stand and testified that AI capable of surpassing human intelligence could arrive within the next year. He reiterated his criticism of OpenAI's nonprofit-to-for-profit conversion and is seeking governance changes that could reshape how AI labs transition between corporate forms. Whatever the case's outcome, the testimony will circulate as a primary-source document for years. The institutional response to AI is no longer in the early-debate phase. Courts, agencies, academies, professional associations, and standards bodies are all writing rules at once, often inconsistently. The next year will be about reconciling them — or surviving the friction when they conflict.