Inference Boom Fuels Cerebras IPO
Research Rundown #185, plus a new memo on Apex
Cerebras’s Blockbuster IPO
Cerebras, an AI chip maker that builds what the company describes as “the world’s largest and most powerful processor for AI,” went public this Thursday on the Nasdaq. It raised $5.6 billion at $185 per share at IPO, then rose 68% to hit $311 per share, representing a $95 billion market cap. Its successful IPO was the end of a long and winding road for Cerebras, which first filed to go public in September 2024 before withdrawing the filing after it became clear that 85% of its revenue was concentrated in a single UAE customer, drawing scrutiny. By the time it refiled in April 2026, it had closed a $20 billion multi-year deal with OpenAI and a binding term sheet with AWS to deploy its chips inside Amazon's data centers.
The underlying thesis remains the same as it was when we first covered Cerebras. The company bet early on the idea that AI is a communication-bound problem and built the world's only commercialized wafer-scale processor, 58x larger than Nvidia's B200. Cerebras says its inference solution delivers answers up to 15x faster “than leading GPU solutions” (i.e., Nvidia). Cerebras’s revenue, fueled by the explosion of demand for inference, grew from $290 million in 2024 to $510 million in 2025, and the company went from a $481 million loss to $88 million in net income over the same time period. If the projection that inference will surpass training to become the dominant AI workload by 2030 is accurate, Cerebras may just be getting started.
Google Signs on for Orbital Data Centers
SpaceX and Google are in talks to explore a rocket-launch partnership as both companies pursue orbital data centers. Google, which owns 6.1% of SpaceX and has a board seat, announced its own orbital computing initiative last year called Project Suncatcher, targeting prototype satellites by 2027 in partnership with Planet Labs. SpaceX has separately filed an application to launch up to 1 million satellites to support its orbital data center ambitions, while Anthropic has already expressed interest in partnering with SpaceX on orbital compute as part of its Colossus deal.
The appeal of orbital data centers is that solar-powered satellites eliminate the land and power constraints that make terrestrial data centers increasingly difficult to site and build. Whether the engineering works at scale is a different question, and many industry experts remain skeptical. Nevertheless, the Google talks are the latest in a series of pre-IPO deals that have rewritten SpaceX’s business. As we wrote in our newsletter last week, SpaceX is using its IPO momentum to position itself less as a launch provider and more as the dominant infrastructure layer of the AI era. The orbital data center thesis is the most speculative piece of that. The Google partnership, if it closes, would at minimum signal that the world’s most sophisticated cloud infrastructure company thinks the bet is worth making.
AI-Driven Rare Book Demand
In January, Anthropic’s Project Panama made headlines when the company was sued for using copyrighted works in its model training data without compensating the authors. Though a judge ruled in Anthropic’s favor, determining that using books that the company paid for to gather training data for fair use, there was continued backlash after the release of internal company documents that said “Project Panama is our effort to destructively scan all the books in the world… We don’t want it to be known that we are working on this.”
Recently, online chatter has swirled about increased orders for rare books, ranging from old computer manuals to texts on architecture, art, and history, all shipped to a handful of warehouses, suggesting that Project Panama-like activity may be picking up again. One online book buyer and seller, Zoom Books, may be behind the surge, with Amazon sellers reporting increasing traffic. On a since-deleted webpage, the company said: “Zoom Books operates in the legitimate used book wholesale market and can supply AI training programs with documented purchase trails across all major genres.” Previously, Anthropic purchased books through Better World Books and World of Books UK.
Latest Research
Apex is a satellite manufacturing company that reached a $1 billion valuation last September. Read our new report here.
The Virtual Cell Revolution is our latest deep dive on how AI could be applied to biology to end the drug development industry’s long stagnation.
What We’re Reading
Google’s Threat Intelligence Group reported that it likely thwarted a criminal group’s attempt to use an AI model to identify and weaponize a zero-day 2FA-bypass vulnerability in a widely-used systems-administration tool, the first real-world AI-developed zero-day GTIG has caught; Google said it did not believe Gemini was used in orchestrating the attack.
The FT reported Anthropic is in talks to raise as much as $50 billion at a ~$900 billion pre-money valuation to fund compute against ~$45 billion annualized revenue (up fivefold from $9 billion at end-2024); CFO Krishna Rao deliberately delayed the round until major SpaceX/Colossus, Google/Broadcom, and AWS compute deals were locked in.
Anthropic separated programmatic Claude usage (Agent SDK, GitHub Actions, third-party tools like OpenClaw) from chat subscription limits, effective June 15, moving to a dedicated monthly credit pool roughly mirroring each tier’s price ($20 Pro, $100 Max 5x, $200 Max 20x), billed at API rates.
Anthropic also launched new Claude for Legal plugins and MCP connectors covering document search/review, case-law lookup, deposition prep, and drafting across commercial, privacy, employment, and AI-governance practice areas, with native integrations to DocuSign, Box, and Westlaw; the updates are seen as a challenge to vertical AI companies Harvey and Legora.
Anduril raised a $5 billion Series H led by returning investors Thrive Capital and Andreessen Horowitz, doubling its valuation to $61 billion (from $30.5 billion last June) after revenue doubled to $2.2 billion in 2025.
The US cleared 10 Chinese firms to buy Nvidia’s H200, each capped at 75K chips, but no chips have been delivered yet, with deals stalling after Beijing’s guidance to its companies; Jensen Huang joined Trump’s delegation to Xi mid-trip after being picked up in Alaska, hoping to break the impasse.
SpaceX paused new applications to its Starlink reseller program, citing “an overwhelming number of submissions,” shutting both the standard and Indirect Reselling tracks as enterprise demand from airlines, cruise lines, mining, emergency response, and rural ISPs ran far ahead of SpaceX’s ability to vet new partners.
NV Energy warned Liberty Utilities it will stop wholesale power supply to 49K Lake Tahoe, California, residents after May 2027, citing up to 5.9 GW of new Northern Nevada data-center demand by 2033, leaving Liberty’s customers, who get 75% of their power from NV Energy, scrambling for replacement supply.
A California federal class action lawsuit accused OpenAI of routing private ChatGPT queries, including topic strings, account identifiers, and email addresses, to Meta and Google via embedded ad-tracking pixels without user consent.
At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!