Reading time: 6 minutes
🗞️ In this edition
Meta Told The IRS Its Data Centers Might Fail
An Appeals Court Said No To Fair Use
DeepSeek Handed Huawei The Missing Software
⚡ YOUR MOVE For people building the life they want, now.
In other AI news:
OpenAI Wants Thirty Billion More
Trump Asked The Labs To Police Themselves
Anthropic Put A Price On An Exploit
4 must-try AI tools
Meta told the IRS its data centers might not work out. An appeals court answered the fair use question for AI training, then sealed the reasons. DeepSeek handed Huawei six pieces of software.
Keep reading.
From Our Sponsor:
Interactive emails used to mean a developer and lots of testing.
Now it takes one message to the AI.
Drag a Stripo Widget (scratch card, quiz, poll, or survey) into your email. Then tell it what you want:
"Make the button bigger"
"Use #AAD000 for the text"
"Add a fourth answer"
It writes the code and makes sure it works in every inbox.
Hide a discount in a scratch card, qualify leads with a quiz, collect feedback with a poll, all without sending anyone to a landing page.
Works on Stripo's free plan. No credit card required.
Try Widgets for Free
Source: The Logic
WHAT HAPPENED
Meta claimed about $3.9 billion in federal research credits against its 2025 taxes, up from roughly $700 million in 2023, and its federal tax expense fell from $9.6 billion in 2024 to $2.8 billion in 2025. The position treats AI data center costs as research and experimental expenditure under Section 174A of the One Big Beautiful Bill Act, which allows the cost to be written off immediately rather than spread over years. Roughly $72 billion of capital spending sits behind the claim, including the Hyperion campus in Louisiana, a 5 gigawatt site with more than $50 billion planned. The New York Times reported the filings on Wednesday and said Meta's own accountants had flagged the treatment as legally risky. The Union of Concerned Scientists warned that if the facilities really can fail, local utility customers could end up carrying the cost of infrastructure nobody uses.
WHY IT MATTERS
Every capital spending number reported this year has been read as conviction. This filing puts a second reading on the same money, because the tax treatment only works if the outcome is genuinely uncertain. A company cannot tell investors a data center is a durable asset and tell the IRS it is an experiment that might not work out, and Meta has now come close to doing both inside one calendar year. The gap between those two statements is the thing to price, not the capex. If the classification holds, every hyperscaler has a reason to file the same way, which quietly moves a large slice of the AI buildout onto the public balance sheet.
OUR TAKE
The obvious read is aggressive accounting by a company that can afford the lawyers. We think it is a disclosure, and the giveaway is that Meta's own accountants flagged it. You do not get flagged for a position you are confident in, you get flagged for one that only stands up if the facility failing is a real possibility.
Watch the IRS, not the filing. If Meta's 2025 research credit survives examination without being reduced, and no hyperscaler restates a research credit, by the end of 2027, then the experiment classification is settled and everybody copies it. If even one of them quietly moves the spending back to amortisation, this was a one year trade and the capex numbers were always the honest part.
Source: Digital Trends
It is the first federal appeals ruling on whether copying protected material to train an AI system counts as fair use, and the answer was no. The trial judge, Stephanos Bibas, had found Ross infringed 2,243 Westlaw headnotes, holding that a headnote carries original expression even when it quotes a public domain opinion, and that the use was not transformative because Ross was building a competing legal research tool. Judge Tamika Montgomery-Reeves wrote the appeal opinion. The reasoning is sealed pending redaction proposals, so the only public document is a one page judgment reading affirmed.Meta shields billions in taxes by classifying AI data centers as experimental models
How did today's stories land?
Source: SCMP.com
The set mirrors the toolkit DeepSeek had already published for Nvidia hardware, and it includes an Ascend version of TileLang, a language for writing high performance kernels with native code generation, automatic scheduling and synchronisation. The company posted the updates to the project's GitHub page and to its official WeChat account, and says the goal is an independent and controllable software ecosystem for the chips. Ascend 950 accelerators are the target. The work lands while US export controls keep Nvidia's top parts out of the Chinese market.
⚡️ YOUR MOVE
FOR PEOPLE BUILDING THE LIFE THEY WANT, NOW.
💸 GPT-6.1 Sol lands close to Astra on capability and charges a fifth of its input and output token prices. If you priced your stack against Astra last week, the arithmetic changed underneath you. → Reprice your stack
🧰 Two of this month's open source security releases exist specifically for agents. Prismor enforces policy before an agent is allowed to execute a tool, and ToolHive runs MCP servers inside containers. → Pick one tonight
🍔 A Big Mac runs $5.69 at one Fresno store and $6.89 at another two miles away, and a Connecticut franchisee was handed a recommendation of $18. McDonald's calls the pricing model a tool, not a mandate. → Check the receipt
🧯 Apple had a plan to replace roughly 5,000 AppleCare staff with AI phone and web agents, and has now shelved it indefinitely. The agents are already answering the line. → Read the reversal
OpenAI Wants Thirty Billion More – The round is being marketed at a $1.4 trillion valuation, and it arrives after OpenAI pushed back its stock market debut into a stretch where Oura, SB Energy and NScale have all delayed their own listings.
Trump Asked The Labs To Police Themselves – Mark Zuckerberg, Greg Brockman, Jensen Huang and Elon Musk signed an accord covering third party audits and an independent safety board, which the President called morally binding and which carries no legal force at all.
Anthropic Put A Price On An Exploit – Zhipu's open weight GLM-5.3 scored 50 of 410 on ExploitBench against 56 for Claude Mythos Preview, and researchers chained a disclosed Chrome bug into a working exploit for $20.40 of API spend.
Grammarly: Catches the sentence you would have regretted sending.
Framer: Builds the site you keep describing instead of shipping.
Gen-2 by Runway: Turns a written scene into footage you never had to film.
ElevenLabs: Gives your script a voice without booking a studio.
Three stories about what a company is willing to put in writing. Meta told the IRS its data centers might not work, in the same year it told everybody else they were the future. A court settled the fair use question for AI training and sealed the reasons for another day. DeepSeek published the software and let the chips make their own case. All three claims are public now. The parts that would let you check them are not.
Ask to see the working.
Interested in featuring your services with us? Email us at [email protected]
