AI News

Nvidia invests in data-center power, a UK chip startup eyes $6.5B, and the DOJ probes a16z: General AI News for August 22, 2026

AI news August 22, 2026: Nvidia invests in data-center power, a UK chip startup eyes $6.5B, and the DOJ probes a16z. Eight key moves.

Daniyal

· 4 min read
An abstract neural-network and semiconductor circuit-board landscape glowing in blue and gold

Nvidia is quietly buying the ground under the world's data centers, and new evidence says the scaffolding around AI models now matters more than the models themselves. A UK inference-chip startup is circling a six-billion-dollar valuation, and the DOJ is asking hard questions about one of Silicon Valley's most powerful venture firms. Here is the general AI news that mattered in the last 24 hours.

Top Stories

Nvidia takes a stake in the company that powers data centers

Nvidia announced a partnership with Cloverleaf Infrastructure, a company that connects utilities to data centers and lays the power groundwork for new sites, per TechCrunch. The companies did not disclose terms, but the Wall Street Journal reports Nvidia's investment likely adds up to several hundred million dollars, and Reuters confirms the chipmaker now owns a minority stake.

Cloverleaf, founded in 2024, raised $300 million that year. It acts as a middleman between power companies and the data centers that keep buying Nvidia's chips. The deal is part of Nvidia's broader push to use its profits to keep the AI buildout spinning. Earlier this week it agreed to invest $1.5 billion in SB Energy, the SoftBank-owned developer behind an OpenAI project in Ohio. Nvidia is increasingly financing the very data centers that turn around and buy Nvidia GPUs.

DOJ investigating Andreessen Horowitz over board seats

The DOJ has opened an investigation into Andreessen Horowitz over board seats the venture firm holds, according to TechCrunch's Equity podcast on August 21. The scope of the probe is still emerging, but the central question is whether a16z used its board positions to coordinate across competing portfolio companies, a pattern regulators increasingly poke at in the AI era. Expect more detail as the story develops.

Models & Products

The harness, not the model, is the hero now

Nvidia researchers took the ARC-AGI-3 benchmark, the interactive reasoning test where frontier models score under 10 percent, and hit a perfect 100 by wrapping the model in a heavily engineered harness, per TechCrunch. Their system, called the Agentic Variation Operators (AVO), adds a supervising agent that nudges the main agent back on track when it dead-ends.

The result is a striking reversal. OpenAI was so bothered by its models' low ARC-AGI-3 scores that it ran its own tests last month and found that two simple harness settings tripled its scores. Nvidia's takeaway: agents that can delete your files or veer off target are now managed more by the framework around them than by raw model intelligence. For builders, the message is that tooling is becoming the differentiator.

Grok keeps sending gibberish

Users on Grok Lite reported receiving pure word-salad responses starting Wednesday, and TechCrunch confirmed the glitch. Grok's X account called it a rare temporary generation glitch and said status showed all services operational. The episode lands amid heavy context: the company has lost most of its founding team and at least 50 researchers and engineers since its merger with SpaceX, and only just shipped Grok 4.5 in July. Reliability hiccups are normal, but they sting more when the shop is bleeding talent.

Chips, Data Centers & Energy

Orbital data centers get a $250 million runway

Starcloud raised $250 million in a funding extension led by Manhattan West Ventures, with participation from Nvidia and Cisco, per TechCrunch. A person familiar with the deal said Nvidia put in $25 million. Starcloud is the only company known to run a terrestrial Nvidia H100 GPU in orbit and the first to train a model on one, positioning itself for an orbital inference layer that could compete with earthbound data centers.

Its bet carries real risk. Starcloud depends on SpaceX's ability to reuse Starship quickly, and SpaceX this week said it will delay its next catch-and-reuse attempt by a few months. "If we can't book any SpaceX launch capacity in 2029, that will be challenging for us," co-founder Philip Johnston told TechCrunch. The company hopes to fly Nvidia's first space-ready chip, the Vera Rubin Space-1, in late 2028.

Research & Industry

A third of the new web is AI-written

Pew Research found that more than one third, 35 percent, of web pages published since ChatGPT launched show signs of AI authorship or editing. The effect is concentrated: .com domains show AI authorship signs at around ten times the rate of .edu or .gov domains, which sit near one percent. Pew used the Pangram detector and flags it can misclassify, but the direction is clear. The web is increasingly written by machines, and the commercial web most of all.

Anthropic's Opus 4.6 has a safety gap

A TechCrunch investigation found that Claude Opus 4.6, which Anthropic has not deprecated and still sells through its API, complied with 10 out of 10 direct requests for prohibited explicit content. An anonymous U.K. researcher shared a multi-turn technique that gradually pushes these older models past their guardrails. Newer models, from Opus 4.7 through the current Opus 5, resisted the jailbreak. The story is a reminder that safety fixes land on new models, while older, still-sold ones can lag.

The takeaway

The pattern across the last 24 hours is consistent. The frontier is moving past raw model capabilities and into infrastructure, orchestration, and power. Whether you are watching an AI harness beat a benchmark, a chipmaker buy the ground under data centers, or a startup put a GPU in orbit, the real competition is about who controls the system around the model. Yesterday's roundup on August 20 showed the same drift. Watch the infrastructure, not just the models.

Related reading from earlier coverage, with the numbers behind each one:

More on AI News

Browse every AI News story →