Meta cut its federal tax bill by calling AI data centers experiments
Three stories this hour: the tax mechanics sitting under the AI buildout, a frontier model driving a humanoid through a kitchen it had never seen, and an agent platform that just handed its agents wallets.
Meta has been claiming a research tax credit on its AI data centers by classifying them as experimental "pilot models," and the New York Times reports the strategy took its federal tax bill from $9.6 billion in 2024 down to $2.8 billion in 2025. The mechanism is a credit created in the 1980s to spur innovation: companies can claim a rebate on supplies, but only when those supplies are being tested in an experimental effort rather than running standard business operations. Meta's argument is that the Nvidia chips it buys belong to the experiment — and by the newspaper's count, the credit has climbed from $700 million in 2023 to $2 billion in 2024 to $3.9 billion in 2025.
The numbers are checkable, and they check. Meta's annual report for fiscal 2025 lists current federal tax of $2.82 billion, against $9.57 billion a year earlier, and states that its unrecognized tax benefits are "predominantly accrued for uncertainties with our research tax credits" — the company's own disclosure that the position is exposed. Andre Shevchuck, a partner at BPM who works on research credits, told the Times the approach was "kind of wild and out there."
The wider point is that this is how the public pays for the buildout — through the tax code, not through a vote. Senator Elizabeth Warren's letter to the companies last week argued the AI tax breaks deserve scrutiny, and the reporting is specific to Meta, so treat the classification as a Meta finding rather than an industry habit. Read it as the subsidy half of the fight that has been running all month over data-center power and water: the earlier round was about buildings, this one is about the silicon inside them.
Stanford and Caltech researchers put a Unitree G1 humanoid into a kitchen it had never seen, with GPT-6 Astra as the planner, and it tidied the room and fetched medicine from a drawer it could not see — without training a new task policy. The robot first walks the space, recording camera frames, lidar returns and its own joint poses; Astra then acts as a Real2Sim agent and rebuilds the kitchen inside Nvidia's Isaac Sim, so the system knows where an object lives and how to walk back to it. From there the model works by tool call: it picks a skill — navigate, pick, place, open the drawer, take from the drawer — instead of emitting joint angles, and a pretrained whole-body controller handles the walking and reaching.
That middle layer is the interesting part. The common pattern puts a learned vision-language-action model between the planner and the motors; this project routes the model straight into a reusable skill library and skips it, which is why a room-sized task needed no new training data. We covered the previous milestone in this line this month — GPT-6 Astra drove a real car through a cone course — and the direction is the same: the model moving outward from the desk.
The limits are still the limits. Perception, planning and skills ran on a laptop with an RTX 4090, Astra was called remotely, rebuilding the digital twin costs setup time and API spend, and the project's own notes list finger-servo overheating during long runs. "No new training" also does not mean nothing was trained — the perception stack and the controller are pretrained prerequisites.
Manus released Manus 2.0 and a new personal agent called Cue, which gives each agent its own email address, phone number, wallet and cloud computer — and lets you put several of them in one group chat to split a job. The release landed Monday, four weeks after the company resumed independent operations, and it puts Manus against Meta's Muse, which had topped the US App Store free chart ten days earlier. Cue is invite-only for now, and Manus says a China-market product is being staffed.
The mechanics are the story. An agent that can send messages, answer your calls and summarise them, and pay within a budget you set is an agent with an identity on the network — the same step OpenAI took when Dots gave Pro users an agent with its own computer. Manus also turned its desktop app into a Studio with a hands-on video timeline and game templates, and says its new Cascade framework cuts token use 23.2%, task time 28.2% and running cost 32%. Those are the company's own figures from one undisclosed configuration, so treat them as claims rather than measurements. The open question is the one every agent vendor now faces: an assistant holding a wallet and a phone number is also a target, and nobody has settled who is liable when it pays the wrong person.
What to watch: whether another hyperscaler gets asked about the same research-credit classification in its next filing, and whether the skill-library approach survives a second, messier house.
If an agent with your wallet spends on the wrong thing, who should be on the hook — the lab, the operator, or you? Tell us in the comments.
Sources: New York Times · Meta fiscal 2025 annual report (SEC) · CNBC · Stanford TML · The Decoder · QbitAI · Manus · Bloomberg · The Next Web
