☕️ Google to pay SpaceX $920M monthly for compute

Hi there, this is your daily ☕️ Techpresso.

In today's Techpresso:

💸 Google to pay SpaceX $920M monthly for compute

🔒 OpenAI rolls out Lockdown Mode against prompt injections

🏛️ Trump administration eyes stake in OpenAI

⛺ Meta builds AI data centers in tents

🛡️ Anthropic is helping the NSA

Plus: 🎁 9 other news you might like, 🧰 6 tools, and 📚 6 papers.

💸 Google to pay SpaceX $920M monthly for compute LINK

Google has signed a deal to pay SpaceX $920 million each month from October 2026 through June 2029 for access to around 110,000 NVIDIA GPUs, CPUs, memory, and related components.

The arrangement mirrors SpaceX's late May deal with Anthropic, which pays $1.25 billion monthly for all compute at the Colossus 1 data center near Memphis, though SpaceX didn't say which site Google would use.

Either side can cancel with 90 days notice after December 31, 2026, and if SpaceX misses the September 30, 2026 GPU delivery target, Google can terminate after a one-month grace period or accept fewer GPUs at reduced fees.

🔒 OpenAI rolls out Lockdown Mode against prompt injections LINK

OpenAI is rolling out Lockdown Mode, an optional security setting in ChatGPT that gives users stricter protection against prompt injection attacks, where bad actors hide malicious instructions on webpages to trick AI systems into leaking data.

Turning it on disables Deep Research and Agent Mode entirely, blocks ChatGPT from pulling or displaying images from the internet, and stops the chatbot from downloading files, though manual uploads and image generation still work.

The feature is available to all personal accounts, including the free tier, and works by limiting network requests an attacker could exploit to exfiltrate sensitive data, rather than stopping injections from appearing in processed content.

🏛️ Trump administration eyes stake in OpenAI LINK

The Trump administration and OpenAI CEO Sam Altman are in active talks about giving the U.S. government an equity stake in the AI company, a conversation that has been going on for more than a year.

Under the potential deal, OpenAI would donate equity to seed a "Public Wealth Fund," an idea from the company's April policy proposal that would invest in long-term assets and share returns with citizens.

No investment terms have been settled, but Trump told reporters on Air Force One that "pieces could be given to the American public," and he said he plans to meet with AI companies in the near future.

⛺ Meta builds AI data centers in tents LINK

Meta is racing to expand its AI computing footprint by housing six data centers inside large tents outside New Albany, Ohio, swapping its original building plans for what it calls "rapid deployment structures."

City permits reviewed by Cleanview's Michael Thomas show Meta put up five 125,000 square foot tents between April and June, a sharp contrast to the two to three years it took to construct the site's first five buildings.

Meta has also signed a decade-long deal for off-grid power plants to run the Ohio tents, with construction nearly finished, and the company is putting up more data center tents at its AI site in Tennessee.

🛡️ Anthropic is helping the NSA LINK

Anthropic has embedded around six engineers at the National Security Agency to deploy Mythos, its most capable AI model, for offensive cyber operations, according to a Financial Times report published Thursday.

Mythos is the same model Anthropic has refused to release publicly over misuse risk, restricting it to vetted partners through Project Glasswing, a coalition that includes Microsoft, Apple, and Amazon.

The NSA kept using Mythos even as Anthropic sues the Pentagon over a supply-chain risk label from Defense Secretary Pete Hegseth, which followed a collapsed $200 million contract over autonomous weapons and domestic surveillance limits.

Other news you might like

• New York legislature passes 1-year data center moratoriumLINK

• Kalshi asks paid influencers to delete posts sowing doubts over LA mayoral electionLINK

• Chrome for Mac breaks benchmark records on the latest MacBook ProLINK

• Meta's Hatch AI agent could cost up to $200 a month and marks its first paid AI productLINK

• LinkedIn co-founder Reid Hoffman to leave Microsoft board after nearly a decadeLINK

• Former cyber executive turned whistleblower accuses IBM of covering up several data breachesLINK

• Meta's stock sinks on report company could raise tens of billions of dollars to fund AI pushLINK

• Elon Musk's xAI reportedly trained its coding models on Claude outputs for months before getting cut offLINK

🧰 Trending tools

Google Search Profiles: lets you switch between distinct search identities, keeping personal and work-related queries, history, and preferences cleanly separated.LINK

Agent Mode on Arena: lets you benchmark autonomous AI agents across multiple frontier models handling real multi-step tasks like browsing, research, and coding simultaneously.LINK

Manus Shopify Connector: lets merchants build, manage, and market a Shopify store entirely through natural language conversation, eliminating manual setup steps.LINK

MAI-Image-2.5: a production-ready image generation and editing model with precise localized edits, text rendering, and identity preservation built in.LINK

QWERTYS: a Tetris-style typing puzzle where you drop letter blocks onto correct keyboard positions, testing your layout memory under pressure.LINK

Gaming services by IFTTT: connect games and platforms to automate stream announcements, track milestones, and get deal alerts without manual effort.LINK

📚 Trending papers & reports

AI math skills now solve all but 2 out of 100 expert-level math questions created by 49 mathematicians, showing these tools are getting remarkably close to research-grade mathematical reasoning.LINK

Decision trees and AI generators are secretly the same math underneath, letting a new method match top performance within 2% on many benchmarks while running twice as fast on structured data tasks.LINK

AI text generation trained with a new fine-tuning method called Discrete Tilt Matching gets noticeably better at logic puzzles like Sudoku and Countdown without losing ground on standard math benchmarks.LINK

AI memory efficiency gets up to 2.6x better throughput in multi-turn conversations with a new system called Tangram, letting AI servers handle more users without sacrificing accuracy.LINK

AI coding agents writing their own tests during bug fixes turns out to be mostly habit, not help, as changing how often they write tests does not significantly improve results.LINK

AI evaluation tools that mimic diverse human perspectives break down over time, becoming inconsistent, so static one-size-fits-all testing methods are not reliable enough to judge whether AI truly aligns with real human values.LINK

Want to get the latest news differently? Find us on:

See you tomorrow for a new dose of ☕️ Techpresso!