Volantis raises $88M to smash the memory wall with lasers
Three moves on the board this afternoon: a photonics startup claims it can blow past the memory wall, Anthropic's government business just went GA, and the open MoE training stack got a trillion-parameter upgrade.
Volantis raised $88 million to move data between AI and memory chips with lasers. The San Francisco startup is using vertical-cavity surface-emitting lasers — the same tiny components that power Face ID on iPhones — to shuttle data between a GPU and its memory optically instead of over copper wires, and according to Reuters it closed an $88 million Series A co-led by Lachy Groom and Abstract Ventures, with John Doerr among the backers. The pitch is blunt: today's Nvidia GPUs sit under eight high-bandwidth memory chips each, while Volantis says its optical link lets a single GPU be surrounded by 220 memory chips, with capacity and bandwidth up roughly 100-fold at under a picojoule per bit. The money and the names are real; the numbers are the company's own, and no valuation was disclosed. What to watch is cadence, not claims — first engines are due to reach customers in 2027, and the memory wall is the single bottleneck every frontier lab is currently paying to route around. If photonics in the memory path works half as well as it does on chip-to-chip links, this is one of the more consequential hardware bets of the year.
Claude for Government is now generally available for US federal and state agencies. Anthropic announced the GA on September 30 — public beta since July — running in a FedRAMP High environment, the strictest US cloud security tier, with no seat fees: agencies buy usage in fixed increments under a hard not-to-exceed cap, with audit logs, per-department budget controls, and two-person approval for sensitive operations. Conversation history stays on the agency's own device. The timing matters more than the feature list: the Pentagon banned Claude in March after Anthropic refused to drop its bans on mass surveillance and autonomous weapons, a San Francisco judge struck down one designation as unlawful retaliation in August, a Washington appeals court upheld the other in late September, and Dario Amodei dined with Trump on Sunday. Selling to the civilians while the fight with the military drags on is exactly the split strategy the court rulings have left Anthropic with — and it suggests the company has decided federal civilian business is worth building out regardless of how the defense saga resolves.
Ai2 open-sourced Olmo-core 3, a training stack built for trillion-parameter MoEs. The Allen Institute for AI's upgrade to the framework behind Olmo redesigned its MoE training system around how experts actually live on GPUs, and the benchmark that matters: growing the expert pool from 8 to 128 while still activating only four experts per token cut training throughput by under 5 percent, and on eight Nvidia B300 GPUs the new stack hit 52,000 tokens per second per GPU against 19,400 for the old implementation — about 2.7 times faster. Ai2 says the same infrastructure has been benchmarked at 1.2 trillion total parameters across 512 GPUs, and its next-generation Olmo will be an MoE trained on it. There's also honest negative results in the tech report — a routing-balance score that improved while the actual workload got less balanced, which the authors call token gerrymandering. Open training infrastructure rarely makes headlines next to model releases, but it's the part of the stack academic labs can actually afford to copy.
What to watch: whether Volantis's 2027 customer timeline survives contact with a foundry, and whether Anthropic's civilian-first strategy draws a response from the Pentagon.
Which of these would you put money on — the photonics bet, the government wedge, or the open stack? Tell us in the comments.
Sources: OLMo-core (GitHub)
