Meta is negotiating to rent its AI rival Anthropic up to 10 billion dollars in computing power, proof that chips now matter more than allegiance. Alibaba previewed a 2.4-trillion-parameter model it plans to open up, betting open weights beat closed rivals on cost and control. Netflix confirmed generative AI touched roughly 300 titles this year, cutting some production timelines in half. Here's what founders need to know today.

Meta is negotiating to lease its AI computing power to rival Anthropic in a deal that could reach 10 billion dollars over two years, with Anthropic paying in monthly installments and either side able to walk away before it concludes. The talks reportedly started in June and are still early, with no signed agreement yet. It would be smaller than Anthropic's 45 billion dollar, three-year compute deal with SpaceX signed in May, but it marks the first time Meta would be renting capacity to a direct AI rival. Both companies declined to comment officially.

Why it matters: Compute has become as strategic as the models themselves. Meta is turning spare infrastructure into a cloud business, and Anthropic is spreading its chip supply across multiple partners rather than depending on any single one, including former partners in unrelated industries like SpaceX.

What to do:

Source: Yahoo Finance

Alibaba unveiled Qwen3.8-Max-Preview, a 2.4-trillion-parameter multimodal model it plans to release with open weights, two days after Moonshot AI shipped its own 2.8-trillion-parameter open model, Kimi K3. Alibaba claims the model is "second only to Fable 5" among systems it benchmarked, but has not published an official benchmark table or model card, and no independent testing exists yet. Developers can test it now through Alibaba's Token Plan at 10% of standard pricing, though Alibaba itself advises against production use until benchmarks land.

Why it matters: If the performance claims hold up under independent scrutiny, businesses could run a highly capable model in-house instead of paying for closed US models, trading a subscription cost for a hosting and engineering one.

What to do:

Source: MarkTechPost

Netflix co-CEO Ted Sarandos said on the company's Q2 2026 earnings call that creators used generative AI workflows across around 300 titles this year, including 17 minutes of AI-enhanced footage in the documentary series American Experiment. Sarandos said those AI-assisted sequences were produced twice as fast and at half the cost of traditional methods, crediting internal teams and Netflix's recent InterPositive acquisition. The disclosure signals AI has moved from experiment to a standard line item in the production budget.

Why it matters: When a major studio puts a number and a cost-and-speed multiple on AI's impact, it changes the baseline every production vendor and post-house gets measured against.

What to do:

Source: MediaNama

URKL (Ultimate Robot Knock-out Legend), billed as the world's first humanoid robot free-combat league, held its opening night in Shenzhen on July 17. During the main event, a robot named White Eagle landed a kick that knocked the head clean off its opponent, Matador, which kept fighting briefly with its head dangling before collapsing. The tournament, organized by EngineAI, features 32 international teams competing with T800 humanoid robots for a $1.44 million prize.

Why it's interesting: It is somehow both slapstick and a genuine showcase of how durable and fast humanoid robot hardware has gotten in a single year.

Key takeaway: Combat sports are becoming an unofficial stress test for humanoid robotics, and the hardware is holding up better than expected, headless or not.

Source: Newsweek

A designer at Mixfont released Decoy Font, a typeface that uses layered spatial frequencies to show one message to humans reading normally and a different, misleading one to AI systems analyzing the image closely. Early testing reportedly fooled several major chatbots reading screenshots of text set in the font. It spread quickly across Hacker News and tech press as a working proof of concept for keeping content readable by people while resistant to AI scraping.

Why it's interesting: It flips the usual AI-versus-humans framing. This is a tool built specifically to make text illegible to machines while staying legible to people.

Key takeaway: Expect more "AI-resistant" design tools as more people look for ways to opt out of being scraped, at least until the models catch up.

Source: Mixfont

OpenAI detailed GPT-Red, an internal automated red-teaming model built to attack OpenAI's own systems and surface vulnerabilities before release. In head-to-head testing, GPT-Red beat human red-teamers 84% to 13% at finding prompt injection weaknesses, and OpenAI credits it with making GPT-5.6 more resistant to that attack type. It runs continuously against new models rather than as a one-time pre-launch check.

Why it's interesting: The idea of a company building a dedicated AI whose only job is to break its own AI is a very literal version of "fight fire with fire."

Key takeaway: Automated red-teaming is quietly becoming standard infrastructure at frontier labs, not just a research curiosity.

Source: MarkTechPost

Context: With compute now being rationed, rented, and renegotiated between rivals like Meta and Anthropic, founders need a fast way to pressure-test their own infrastructure dependencies before a vendor or partner situation changes under them.

Prompt: You are my infrastructure risk advisor. Context: [paste your current stack, key vendors, and what each one provides]. Task: identify the single most concentrated point of failure in this setup, explain what would break if that vendor changed terms or became unavailable, and propose one realistic mitigation I could implement within 30 days.

Do not: Do not recommend switching vendors outright without first identifying the actual failure mode.

If/Then: If the analysis surfaces a single point of failure with no backup, then treat sourcing a second option as a priority task, not a someday item.

Example: A founder running their product on a single AI API pasted their stack notes and learned their entire onboarding flow depended on one vendor's uptime with no fallback, prompting them to add a second provider as a failover within the month.

U.S. District Judge William Orrick declined to block Meta from finalizing layoffs affecting 26 employees who sued the company, saying they hadn't shown the "irreparable harm" required for an emergency order. The employees allege Meta used AI performance scoring that penalized workers on medical, parental, or bereavement leave; Meta says the decisions were made by humans, not AI. A separate motion for a longer-term injunction is still pending, and the judge signaled he may reconsider based on further evidence about how the AI was actually used. Many of the affected layoffs are scheduled to finalize on July 22.

Why it's important: This is one of the first real courtroom tests of whether AI-assisted workforce decisions can create legal liability distinct from ordinary layoffs, and the ruling so far favors the employer.

Business takeaway: If you use any algorithmic or AI-assisted scoring in HR decisions, document how humans reviewed and can override it before you need that record in a dispute.

Source: Insurance Journal

Anthropic CEO Dario Amodei donated 1 million dollars of his own money to Public First Action, a super PAC pushing for mandatory pre-deployment testing of frontier AI models and regulatory power to block dangerous systems; Anthropic employees added over 1 million more, and the company itself pledged 20 million dollars to the group earlier this year. On the other side, Leading the Future, backed by OpenAI co-founder Greg Brockman, is reportedly planning to spend up to 125 million dollars pushing a lighter-touch regulatory approach. Total AI-related political spending has already topped 50 million dollars ahead of the midterms.

Why it's important: AI labs are no longer just lobbying regulators, they're funding the campaigns that decide who the regulators are, and the two biggest labs are backing opposite philosophies.

Business takeaway: Whichever side wins more races in the midterms will likely shape the compliance environment your AI vendors operate under within the next two years, so it's worth tracking now rather than after rules land.

Source: Crypto Briefing

Roughly 35,000 Hyundai workers began a rolling strike after the company confirmed plans to deploy more than 25,000 Boston Dynamics Atlas humanoid robots across its Hyundai and Kia plants, with production costs per robot expected to fall from about 135,000 dollars today to roughly 30,000 dollars at scale by 2028. The union's core demand is that no robot enters a Hyundai workplace without a labor-management agreement first, alongside separate wage and profit-sharing issues. No deployment date has been set for the Korean plants affected, which produce roughly half of Hyundai's global output.

Why it's important: This is reportedly the first labor strike specifically triggered by humanoid robot deployment, and it's happening well before the robots have even arrived on the floor.

Business takeaway: If your business plans to introduce robotics or AI-driven automation into a unionized or people-heavy workflow, negotiate the terms of that rollout before deployment, not after.

Source: Forbes

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas