SpaceX IPO beckons and OpenAI's Wall Street waltz

SpaceX shoots big. SpaceX’s long-awaited IPO filing reads less like a standard S-1 and more like Elon Musk’s master plan in spreadsheet form. The company is still powered by Starlink, which brought in about $11 billion last year, but it is also pouring huge money into AI, with roughly 60% of 2025 capital spending, around $20 billion, going to the xAI division. That bet is expensive, and so is Starship, which remains the hinge for nearly everything ahead, from launching next-gen satellites in 2026 to chasing Mars, orbital data centers, and even point-to-point Earth travel. The numbers are eye-popping: $18 billion in revenue, a $4.9 billion loss in 2025, and more than $37 billion burned since inception. Through it all, one thing is crystal clear: this is still Musk’s company, with majority voting control and a compensation package tied to Mars colonies and space compute. More here.

Rekursor boosts a benchmark. Quick one from my own company. We just put Rekursor's continual learning pitch to the test on Harvey LAB, Harvey AI's open legal benchmark. On a corporate governance task with 48 rubric criteria, Harvey's own agent setup, running Claude Sonnet 4.6, scored 45 out of 48, which still counts as a fail under Harvey's all-pass grading. We took that first run, turned the gaps into two reusable "skills," and reran the exact same task with no model changes, no fine-tuning, and no harness edits. Result: 48 out of 48, with zero regressions. The missing pieces weren't basic mistakes, they were legal judgment gaps around conflict of interest and Delaware safe harbor rules. This is one task, but we plan to sweep more tasks next. The plan is to keep stacking results like this and make AI agents smarter over time. More here.

OpenAI eyes Wall Street. OpenAI may be getting ready for its biggest plot twist yet: going public. The ChatGPT maker is reportedly working with bankers on a confidential IPO filing that could land in the coming days or weeks, with some sources saying as soon as Friday. Heavyweights like Goldman Sachs and Morgan Stanley are said to be helping shape the draft prospectus as OpenAI lines up what could become one of the most closely watched tech listings in years. If it happens, the move would mark a major new chapter for Sam Altman’s AI giant, which has already become one of the defining companies of the generative AI boom. For investors, rivals, and basically anyone watching the AI race, this is the kind of filing that could reset the conversation fast. More here.

Compute cash cannon. Anthropic is shelling out a jaw-dropping $1.25 billion a month to SpaceX through May 2029, putting the annual tab at roughly $15 billion for AI compute. That is a huge number for any company, especially since SpaceX pulls in about $18 billion a year total. The deal, first announced without a price tag, got new detail in SpaceX's IPO filing, which also noted slightly lower payments in May and June as things ramp up. Anthropic, meanwhile, says it is already expanding from SpaceX's Colossus 1 to Colossus 2, adding more Nvidia GB200 capacity as it races to keep up with surging demand. SpaceX also made clear this is not a one-off, saying it expects to sign more compute contracts. One catch: either side can walk away with 90 days' notice. More here.

Math gets rocked. OpenAI just dropped a genuine math bombshell: one of its general-purpose reasoning models has disproved a central conjecture in discrete geometry, a problem that has stumped mathematicians since Paul Erdős posed it in 1946. The question sounds innocent enough, how many pairs of points in a plane can sit exactly one unit apart, but it has become one of the field’s most famous headaches. For decades, researchers thought square-grid style constructions were basically optimal, topping out near \(n^{1+o(1)}\). OpenAI’s model found an infinite family of counterexamples with a real polynomial improvement, at least \(n^{1+\delta}\), with later refinement suggesting \(\delta = 0.014\). The proof was checked by outside mathematicians, and heavyweights like Tim Gowers and Noga Alon are calling it a milestone. Big picture, this is less “AI assistant” and more “AI researcher.” More here.

Anthropic hits escape velocity. Anthropic is having a seriously big moment. The AI startup is projecting second-quarter revenue of $10.9 billion, more than double from the prior period, and that surge is expected to push it into operating profitability for the first time. That is a notable flex in an industry where the standard story is sky-high costs, endless chip spending, and profits that always seem just out of reach. The numbers surfaced as part of an ongoing fundraising round, one that could value Anthropic above OpenAI and cement its transformation from perceived AI underdog to top-tier contender. In short, this is not just a growth story, it is a narrative reset. Anthropic is showing investors that an AI company can scale at breakneck speed and still start making real money sooner than expected. More here.

Cohere goes open. Cohere just made a big play for enterprise AI with Command A+, a 218 billion parameter model that is finally being released under a true Apache 2.0 open-source license. That means companies can actually use, modify, and commercialize it without the usual licensing headaches. The real hook, though, is efficiency. Thanks to a sparse MoE design, only 25 billion parameters are active at once, and Cohere says its new 4-bit quantization keeps performance nearly intact while letting the model run on a single Nvidia Blackwell B200 or two H100s. Command A+ is also built for the real-world stuff enterprises care about: agentic workflows, multimodal document processing, a 128K context window, and native citations that tie answers back to source data. Add better token efficiency across 48 languages, and Cohere is clearly aiming at global, regulated, cost-conscious deployments. More here.

Server heat spills. Data centers are not just powering the AI boom, they are also warming the neighborhoods next door. Researchers at Arizona State University found that Phoenix-area facilities can push downwind air temperatures up by 1.3 to 1.6 degrees Fahrenheit on average, with spikes as high as 4 degrees. That extra heat can travel about a third of a mile, roughly five city blocks, from the site. The team says this is the first study to directly measure real-time temperature differences upwind and downwind of data centers, using sensor-equipped cars around four Phoenix facilities. The big issue is scale: a single center can dump more waste heat than 40,000 households, and U.S. data center capacity is expected to more than double by 2030. The upside is researchers think smarter cooling designs, greenbelts, and tighter city permitting could help cool things down. More here.

Pentagon's AI sprint. The Pentagon is moving fast to get powerful AI tools onto some of the government’s most sensitive networks, with a new task force spanning U.S. Cyber Command and the NSA. The mission is pretty straightforward, if not exactly simple: figure out how models from OpenAI, Google, and others can be safely used for cyber operations, including on classified systems packed with top-secret intelligence. The push comes as frontier AI models get alarmingly good at finding software flaws and even helping exploit them, raising fears that elite cyber capabilities could soon trickle down to less sophisticated hackers. Gen. Joshua Rudd reportedly told staff the task force will tap NSA’s technical muscle, while Cyber Command may take the lead operationally. It also lands as the White House weighs tighter oversight of advanced AI, and as Anthropic’s Mythos remains tangled in a legal fight with the Pentagon. More here.