Bloomberg reported on July 1 that Meta is building a cloud business to sell its excess AI compute to outside customers, citing people familiar with the matter. Reuters said it could not independently verify the report, and Meta declined to comment. The effort sits inside "Meta Compute," the internal infrastructure program Mark Zuckerberg launched in January to scale Meta's own buildout toward "tens of gigawatts this decade, hundreds of gigawatts over time." At the May 27 shareholder meeting, Zuckerberg said an external compute business was "definitely on the table," adding that outside companies approach Meta "almost every week" to buy spare capacity at a premium. Two models are reportedly under consideration: hosting Meta's own models for external developers via API, akin to AWS Bedrock, or renting bare-metal GPU capacity directly against neoclouds like CoreWeave and Nebius.
No launch date, pricing, or capacity figures have been disclosed.
The US Commerce Department lifted export controls on Anthropic's Fable 5 and Mythos 5 models on June 30, ending an 18-day standoff that began three days after their June 12 launch. A jailbreak flagged by a "trusted partner," reported to be Amazon, triggered the freeze; Anthropic says a new safeguard now blocks the exploit 99% of the time. Access is restoring across AWS, Google Cloud, and Microsoft Foundry — the first case of Commerce gating a frontier model's availability over a safety defect rather than a China-specific hardware restriction.
Tenstorrent launched TT-Ascalon S, a RISC-V CPU core for agentic AI workloads, at a Tokyo event tied to a Japan sovereign-compute buildout: more than 120 Galaxy systems deployed with named partners Rapidus, Preferred Networks, Socionext, and Turing. The company claims 900 tokens per second per user running Kimi K2.6 on Galaxy Blackhole, three times GPU alternatives. CEO Jim Keller also denied June reports of Qualcomm acquisition talks, saying he's discussing IP licensing with both Qualcomm and Intel but isn't selling.
Korean NPU maker Rebellions acquired AI-inference optimization startup SqueezeBits in an all-stock deal, making it a wholly owned subsidiary. SqueezeBits posted roughly $1.9M in revenue for fiscal 2025 with an operating margin near 34%, and has prior work alongside both Intel and Nvidia. Kakao Ventures, an investor in both companies, receives Rebellions shares in the swap. CEO Park Sung-hyun called it a move toward a full-stack position spanning NPU hardware, inference software, and serving — ahead of a planned 2026 IPO following March's $400M pre-IPO round at a $2.3B valuation.
Montreal-based Stathera raised an oversubscribed $55M Series B, led by new investor Maverick Silicon with Celesta Capital, BDC Capital, MediaTek Innovation Fund, TXC Corporation, and Ultratech Capital Partners participating, bringing total funding to $75M. The company makes MEMS-based silicon oscillators that replace quartz timing crystals, with up to 85% smaller footprint and no external load capacitors. Precision clock synchronization is emerging as a distinct AI-datacenter bottleneck alongside power and cooling — jitter and sync errors degrade GPU cluster, switch, and optical-interconnect performance at scale — and the raise signals infrastructure capital rotating into a layer beneath the chip itself.
No issue of The Compute 100 will go out. Publication resumes Monday, July 6.