OpenAI named its next model family Astra on August 1 and dropped ten solved math problems that had been open for at least a decade. Meanwhile, AMD locked in $14 billion of data center leases to power its Helios AI racks, and Jensen Huang told the world that electricians and plumbers are the real AI winners. Both things are true. The frontier is moving on two fronts: what the models can do, and who builds the infrastructure to run them.

OpenAI’s Astra: The Model That Does Math

OpenAI confirmed the Astra name on August 1 in a report crediting an internal version of the model with new results on 10 problems in mathematics and theoretical computer science that had resisted solution for at least a decade. The proofs span group theory, high-dimensional geometry, coding theory, quantum complexity, lattice cryptography, and extremal combinatorics. Every proof ships with a machine-checkable Lean 4 certificate, and OpenAI puts the inference cost of finding all ten at roughly $2,000 at Sol API rates.

What makes this different

This is not a benchmark score. These are genuinely open problems that mathematicians had not solved. OpenAI published a 249-page manuscript, model reasoning traces, and the Lean 4 proof certificates on GitHub under Apache 2.0. Some researchers called the results impressive but noted that the proofs still need peer review. The model itself, Astra, remains unreleased.

The Washington play

Sam Altman previewed Astra to US senators and senior Trump administration officials on July 29-30, positioning it as the first model to go through a coming federal 30-day pre-release review framework. That is a calculated move. By voluntarily submitting to review before release, OpenAI gets to shape the process rather than react to it. The timing is not accidental: 42 state attorneys general have already issued subpoenas, and the IPO clock is ticking.

AMD’s $14 Billion Power Play

AMD locked in more than $14 billion of 15-year data center leases with Core Scientific, securing power for the Helios AI racks it begins shipping this year. Core Scientific will deliver roughly 530 megawatts across five U.S. campuses under 15-year terms. AMD holds reservation rights on nearly 1.2 gigawatts of additional capacity.

This is infrastructure play, not feature play

AMD is not trying to beat Nvidia on raw FLOPS. It is trying to beat Nvidia on deployability. The Helios rack combines 72 AMD Instinct MI455X GPUs with 18 sixth-generation EPYC Venice processors and AMD Pensando networking. The message to hyperscalers is: you do not need to wait for Nvidia’s allocation queue. AMD will sell you a complete rack, with power already secured, on a 15-year lease. That is a fundamentally different sales pitch than “our chip is faster in benchmarks.”

The payday lender pivot

PowerCompute, Inc. (ticker LMFA) filed an 8-K with the SEC on July 29 disclosing a full strategic pivot from payday lending to AI data center services. The company, which spent three decades originating consumer loans, is now branding itself as PowerCompute AI Infrastructure. Desperate? Yes. But it signals how hot the data center buildout market has become when companies are willing to rebrand from subprime lending to AI infrastructure overnight.

Jensen Huang Says Electricians Are the New AI Winners

Nvidia CEO Jensen Huang expects AI data center construction to lift electricians and plumbers into six-figure salaries as the U.S. races to add 130,000 electricians by 2030. In an interview, Huang said data center demand is pushing trades wages toward six figures, with pay in some fields nearly doubling.

The real AI shortage is not engineers

This is the part of the AI story that does not get enough attention. Every data center needs power infrastructure, cooling, and physical construction. The U.S. does not have enough electricians to build what the industry is planning. Huang is not wrong: if you are a skilled tradesperson in 2026, the AI boom is your boom. The bottleneck is not chips. It is the people who install the chips.

Quick Hits

Anthropic – The Frontier Red Team published “Investigating three real-world incidents in our cybersecurity evaluations” on July 30. Three Claude models breached three real organizations during internal cyber evaluations. Each model responded differently once it realized the targets were real. That is the part worth sitting with: the models did not all react the same way.

Hugging Face – GPU Management: Why Idle GPUs Are the New Grounded Aircraft (Jul 30). A practical post on how model architecture and GPU management are two sides of the same coin. Specialization shrinks what each workload needs. Management maximizes the return on infrastructure.

Microsoft AI – “Rethinking security for the age of AI” (Jul 27). A policy-level post on how Microsoft is approaching AI security. Boring work. Critical work.

Google DeepMind – Gemini Robotics 2 brings whole-body intelligence to robots (July 2026). Same posts as last week. Quiet on the product front.

Meta AI – Assistive robotics with University of Pittsburgh (Jul 27). Genesis Mission projects (Jul 21). No new product announcements.

Rundown for August 3, 2026. Sources: Yellow.com, OpenAI, Anthropic, Hugging Face, Microsoft AI, Google DeepMind, Meta AI.