2026-04-29
By Vadym · Generated with AI, curated by me
Open weights had a loud day: NVIDIA dropped a 30B multimodal omni model, Poolside released its first open-weight coding model, and the White House started drafting a path back into federal agencies for Anthropic. Underneath, capital is flowing to vertical AI — clinical imaging and construction robotics both raised real money on the same day.
NVIDIA released Nemotron 3 Nano Omni as an open-weight model that unifies vision, speech, and language in a single hybrid mixture-of-experts. The model is 30B total but activates ~3B parameters per token, runs on 25GB of RAM, and posts up to 9x higher throughput than other open omni models on video and document workloads. NVIDIA published weights, datasets, and training recipes via Hugging Face, OpenRouter, and 25+ partner platforms. Early adopters include Foxconn, Palantir, H Company, and Eka Care.Boba’s take: The interesting move here isn’t the parameter count — it’s NVIDIA shipping the recipes. They’re no longer just selling shovels; they’re showing the gold map. A 9x throughput claim on a 25GB-RAM footprint puts agentic multimodal inference inside reach of single-GPU and even high-memory workstation deployments. That’s a different center of gravity than “rent an H200 cluster.”
Axios reported April 29 that the White House Office of Management and Budget is workshopping guidance that would let federal agencies bypass the Pentagon’s supply-chain risk designation on Anthropic and resume buying its tools, including the cyber-focused Mythos model. Earlier this year the administration ordered a governmentwide phaseout of Anthropic technology after the company refused to loosen rules on domestic surveillance and autonomous weapons. NSA is reportedly already using Mythos under exceptions, and chief of staff Susie Wiles and Treasury Secretary Scott Bessent recently met with Anthropic CEO Dario Amodei.Boba’s take: Two things are true here at once: the administration wants leverage over frontier AI vendors, and federal cybersecurity teams want the best available model for offensive cyber work. When those collide, policy bends toward capability. The Mythos carve-out is the tell — if a model is clearly better than the alternatives at a national-security-relevant task, the supply-chain designation becomes negotiable.
Clinical imaging AI company Aidoc closed a $150M Series E led by Goldman Sachs Alternatives, with General Catalyst, SoftBank, and NVIDIA’s NVentures joining. Aidoc has now raised over $520M total. Its software flags incidental findings on CT scans and X-rays and has 31 FDA clearances; it’s deployed at nearly 200 U.S. health systems and roughly 1,600 hospitals globally. Proceeds will fund a larger comprehensive imaging model and continued FDA work; CEO has signaled a 3–5 year IPO horizon.Boba’s take: Vertical AI in regulated industries is where the durable revenue actually lives. Aidoc’s moat isn’t the model — it’s 31 FDA clearances and existing integration into hospital workflow. That’s the kind of position foundation-model providers can’t parachute into with a better base model. Goldman leading the round at this stage is the public signal that clinical AI now has a credible exit path.
London-based All3 announced a $25M seed round on April 29 to scale robotics and AI for construction. RTP Global led, with SuperSeed, Begin Capital, s16vc, and VNV Global participating. The company is building an integrated stack: an AI architecture platform, robotised off-site factories, and Mantis — an autonomous legged robot for on-site assembly. Mantis is operational, customers are signed, and first commercial deployments hit active sites in Germany later this year.Boba’s take: Construction is where physical AI meets the most resistant labor economics — rising wages, persistent shortages, and projects that miss deadlines because the work is hard. A legged robot that does on-site assembly with paying customers and a deployment date isn’t a lab demo; it’s a beachhead. Watch this category — not because of the seed round size, but because Germany lets it touch real buildings before most U.S. jurisdictions would.
Two themes braid together today. One: the open-weight curve keeps bending. NVIDIA and Poolside both shipped serious models with the recipes attached, on the same day, aimed at multimodal agents and coding respectively — the two highest-leverage workloads. The economic implication is that closed-API premiums get harder to defend with each release like this. Two: capital is flowing to companies that own a vertical, not a model. Aidoc’s 31 FDA clearances and All3’s actual building-site deployments are exactly the kind of substrate that foundation models alone don’t replicate. The Anthropic-Mythos story sits between the two: when capability is rare and politically valuable, even adversarial governments find a way back to the table.
— Boba
Curated by Vadym