Modal crosses $1B run rate, raises $750M at $15.75B

Good morning 👋 A serverless-GPU company just posted a revenue number most foundation model labs can't touch, and the market priced it accordingly.

In today's issue:

• 🔭 Modal crosses $1B run rate, raises $750M at $15.75B
• 🔬 An elastic transformer that resizes at inference, bias benchmarks that lie once you translate them, and research agents that do better when they're forced to forget
• 🚀 Salience Labs raises $125M to wire GPUs with light instead of copper
• 🏛️ DeepMind publishes its own Gemini 4 Argon safety report, unprompted
• 🛠️ Self-host a financial advisor without your balance sheet leaving your network
• 🔗 Anthropic's Q3 number, a trillion-token-a-day agent swarm, and weather sims at 1/300th the cost

🔭 THE ONE THING

🚀 Modal crosses $1B run rate, raises $750M at $15.75B

Modal says it just crossed $1 billion in annualized revenue, and the market backed that up with a $750 million Series D led by Accel, joined by Kleiner Perkins, at a $15.75 billion valuation. Five months ago the company was worth $4.65 billion. CEO Erik Bernhardsson put the growth at "more than 15x year-over-year," with usage and revenue both roughly tripling since the last round. That's a serverless-GPU shop, not a model lab, clearing a bar most foundation model startups still haven't hit. Infra is where the AI money is actually landing right now, not just the chat apps sitting on top of it.

🔬 RESEARCH HIGHLIGHTS

• "Your Transformer Is Secretly Elastic" trains one model, then slices it into whatever size the hardware needs at inference, no retraining. Rank every MLP neuron and attention head by importance once, and you can prune to any sparsity target on the fly. On a DINOv3 ViT-H+/16, cutting 427M parameters at 50% sparsity costs 0.8 points of ImageNet accuracy; push to 60% and you get 1.8x the throughput on an A100 for 1.26x less latency. Tested across 27 encoders from 22M to 8B parameters. If this generalizes past vision backbones, the S/M/L model menu starts looking like a workaround for a problem that's solved. paper

• Bias benchmarks lie to you the moment you translate them. Audit ten open models across eight languages and all 70 model-language pairs score "less biased" once the stereotype test leaves English. Here's the tell: scores on translated items barely correlate with the English originals (0.23 to 0.27), while just rewording the same English prompt correlates at 0.48. That's measurement noise wearing a fairness costume, not real improvement. Translation quality moves the number some, but even faithful translations show nearly the same drop. The authors also flag three specific pipeline bugs that quietly erase bias signal. Anyone citing a lower non-English score as proof their model got fairer is probably reading static. paper

• Automated research agents get worse the longer they remember. Across four tasks, heuristic programming through kernel optimization and model training, longer single-agent runs, more parallel agents, and even agents that talk to each other all plateau the same way: they accumulate history and anchor to their own past attempts. Forcing periodic amnesia (wipe the search state, restart from the best candidate found so far) beats all three: 15% better than parallel agents with full memory, 39% better than one agent grinding alone. Compute wasn't the bottleneck here. What the agent was allowed to remember was. paper

🚀 AI STARTUPS

• Salience Labs The Oxford photonics spinout closed an oversubscribed $125M Series B to connect GPUs with light instead of copper, led by Bullhound Capital with AMD Ventures and Intel Capital back for seconds. It just demoed a 64-port all-optical switch at ECOC 2026 and is opening a second US office in Silicon Valley. Investor read: Dell'Oro pegs cumulative datacenter switch spending near $1 trillion through 2030. Salience is betting that money lands on scale-up fabric inside the rack, not the links between racks. Announcement

🏛️ POLICY & REGULATION

• Google DeepMind published the full safety report behind Gemini 4 Argon: 12 of 12 simulated cyberattack scenarios solved (the prior model solved 5), external bio red-teaming capped at under 1% of dangerous sub-tasks unlocked, and a three-layer runtime monitor that reads the model's own chain of thought for covert-action attempts. Nobody made them release this. It's DeepMind grading its own homework against its Frontier Safety Framework, and it's the document regulators will reach for the next time someone asks what "safety testing" is supposed to look like. Report

🛠️ TRY THIS

Self-host your own AI financial advisor, keep the heavy inference off your laptop

1. Clone Aurelio Finance (github.com/LosaLosSantos/aurelio-finance, 3,372 stars): tracks net worth, ETFs, cash, and debts, with an AI advisor built in, all running on your own machine.

2. Point the advisor at a model endpoint you control instead of the default. Your account balances shouldn't leave your network to get a recommendation.

3. If your laptop can't run that model fast enough, rent the GPU for the thirty seconds the advisor call needs on Modal, the same serverless infra that just crossed $1B ARR on exactly this pattern. Spin up, answer, spin down to zero.

4. Block outbound traffic except that one inference endpoint. A finance app with an open firewall defeats the point of self-hosting it.

Prompt: Review this net worth snapshot (assets, debts, cash, ETF positions). Flag any single position above 20% of net worth, any debt over 15% APR, and summarize the 30-day cash flow trend in three bullets. No trade recommendations.

🔗 QUICK LINKS

• Anthropic booked $16.8B in Q3, a 45% jump from last quarter, and says it's cash-flow positive ahead of an IPO chasing a $2 trillion valuation. link
• Third Moment Research runs 10,000 agents at once on 36-hour quant research sessions built on OpenAI's GPT-6 Sol and Codex, burning a trillion tokens a day. One of those agents found a caching fix that cut backtest time 97%. link
• AllenAI's HiRO-ACEv2 simulates extreme weather at 3km resolution for about 1/300th the cost of traditional models like X-SHiELD. link

Infra keeps winning quietly this week. Catch you tomorrow.

Pradeep Perugu