The Omniscient Bulletin · 2026-06-24
The Omniscient Bulletin — June 24, 2026
With no new flagship model, the day's AI story was all silicon and capital
Tuesday produced no headline model, and the industry barely noticed, because the action had moved to the infrastructure underneath. China seized the world's supercomputing crown with a machine that uses no GPUs at all, a rebuke to the export controls meant to hold it back. The money told the same story: a cloud ordered $4.1 billion of Nvidia chips, Groq raised $650 million to rent out inference, and a quarter of a billion dollars went to an AI that searches government records. Even the lone model release was about giving capability away. And while everyone else poured concrete, Masayoshi Son told shareholders the AI war will be won on the ground, not in orbit.
Compute
China's new CPU-only supercomputer takes the world's No. 1 spot from the United States
The twice-yearly TOP500 list, released Tuesday at the ISC supercomputing conference, has a new leader: LineShine, a Chinese system at the National Supercomputing Centre in Shenzhen. It reached 2.198 exaflops, pushing the United States' El Capitan (1.809) into second and ending an American run at the top. The twist is how. LineShine uses no GPUs at all, just 13.79 million cores of homegrown 304-core processors. It is the first Chinese machine to lead since 2017, and a sign that export controls have pushed China to win on silicon of its own.
Compute
Argentum signs a $4.1 billion cloud deal for 27,000 of Nvidia's newest GPUs
The scale of AI compute commitments keeps climbing. Argentum, a GPU cloud operating in more than 15 countries, said Tuesday it had signed a $4.1 billion contract to supply roughly 27,000 of Nvidia's newest GB300 chips to an unnamed 'leading AI company,' with capacity arriving in phases starting this year. It follows a $2.5 billion data-center deal the firm struck last month with Boosteroid and DL Invest Group. The customer's anonymity is the tell: at these sums, the labs renting compute increasingly do not want rivals to know how much, or how fast, they are buying.
Industry
Groq raises $650 million to become an AI cloud, after selling its chip technology to Nvidia for $20 billion
Six months after Nvidia paid roughly $20 billion to license Groq's LPU chip technology and absorb most of its senior engineers, Groq raised $650 million from existing investors Disruptive and Infinitum to rebuild as an inference cloud. The irony is baked in: the money will fit out Groq's 13 data centers with Nvidia's own LPX system. Rather than compete on silicon, Groq is now betting the scarce resource is speed and availability of inference at scale, not the chips underneath it.
Industry
Peregrine raises $250 million at a $6.8 billion valuation for its government-data AI
Where is the venture money going now? Into applied AI. Peregrine, which fuses siloed government records into one searchable view for police and agencies, raised $250 million at a $6.8 billion valuation, led by Fifth Down Capital and Sequoia. It says more than 400 agencies covering 125 million people use it, from the Super Bowl to eight of eleven World Cup host cities. The company stresses its privacy controls and audit logs, a sign it knows the same tool that helps find a missing child also unsettles civil-liberties watchdogs.
Frontier
Krea releases an open-weights image model that ranks first among independent labs
The day's one real model drop came from Krea, which released open weights for Krea 2 in two flavors: a research base and a distilled 'Turbo' tuned for fast, few-step image generation. Built on a diffusion transformer with a Qwen text encoder, Krea says the model ranks sixth globally on the Artificial Analysis image leaderboard and first among independent labs. The release fits a pattern: smaller labs giving away capable open-weight models and charging only the largest enterprises, squeezing the closed image generators above them.
Compute
SoftBank's Masayoshi Son calls Elon Musk's space data centers a bad bet
Not everyone is sold on the most exotic AI infrastructure. At a SoftBank shareholder meeting Tuesday, Masayoshi Son waved off Elon Musk's vision of data centers in orbit. The pitch is cheaper solar power, but Son argued that power is a small slice of the cost next to chips, launch, and upkeep, and that beaming data from hundreds of kilometers up adds delay. SoftBank, he signaled, will keep building on the ground. 'In the battle for AI,' he said, 'the next few years will be far more important than what might happen a decade or so from now.'