Today, Modal jumped from our infra unicorn tier up into an infra decacorn. The last time we spoke with Modal CTO Akshat was back in July:
Why AI Infra Must Evolve for Agent Experience (AX) — Modal CTO Akshat Bubna
We’ve been running the “Agent Cloud” series, surveying every top inference/compute/cloud provider from Databricks to Daytona, Railway, and earlier pioneers like E2B…
On the final day of the AIE NYC Summit, Akshat shared a wealth of insights into the infrastructure needed for agents to support “millions of concurrent sandboxes, safely and elastically spinning up RDMA clusters for distributed training, training state-of-the-art speculators for inference, building entirely new sandbox runtimes, and more”:
Akshat Bubna (@akshat_b):
In early 2022, Modal’s first “real” user was using us to crunch large-scale weather data. We sheepishly reached out, asking if they’d be willing to pay around $500 to keep running those jobs. (At the time, we thought that was an enormous sum of money!)
Obviously, the numbers are vastly larger today.
But…Modal (@modal):
Today, we’re announcing that Modal has surpassed $1B in ARR and closed a $750M Series D led by Accel at a $15.75B valuation.
Usage has tripled since our Series C, growing over 15x year-over-year.
We are deeply grateful to all our developers…
As Erik noted in his funding announcement post:
We launched Sandboxes in 2023 when we noticed people starting to use LLMs for code generation. Sandbox usage exploded, and three years later, this single product alone has reached a $300M run rate. Use cases span background agents, RL, vibecoding platforms, and beyond. We’ve spun up nearly 3 billion sandboxes to date, scaling up to 1 million concurrently.
We attended their recent (inaugural?) user conference, and Erik’s keynote is a fantastic way to quickly get up to speed:
AI news from October 13 to October 14, 2026. We scanned 12 subreddits, 544 Twitter accounts, and skipped Discord this time. You can search all past issues on the AINews website. A quick reminder: AINews is now a column under Latent Space.
AI Twitter Recap
Gemini 4 Argon Release: Core Capabilities, Rollout Timeline, and Weaker Observability Footprints
- Rollout: Google has kicked off a broader rollout of Gemini 4 Argon, targeting coding, finance, legal, cybersecurity, and scientific domains. Enterprise customers including Deloitte and Box, along with select startups, have already received early access (Google).
- Availability: Argon is now live globally for Google AI Pro and Ultra subscribers (Gemini App) and has landed on AI Studio alongside a suite of companion demo apps (AI Studio).
- Science access: 10,000 scientists will receive six months of free access to Antigravity’s top tier (GDM).
- Claimed results: DeepMind highlighted breakthroughs across quantum computing, nuclear fusion, and protein design (Thread).
- Enzyme design: Argon reportedly shrank enzyme engineering cycles down to hours, producing 5 lab-validated serine esterases (Details).
- Robotics demo: Using solely video footage of humans playing the game, Argon modeled the underlying physics, wrote its own practice simulation, and taught Apptronik’s Apollo humanoid robot to land 7 consecutive cornhole toss shots (Google).
- Efficiency claims: A Googler shared that Argon outperforms Astra on vision tasks while running at a quarter of the cost (@gabriberton).
- Gemma optimization: Argon helped optimize Gemma 4 E2B on vLLM to over 120K tok/s on a single A100. That marks a roughly 2.9x speedup, hitting 124K tok/s in prefill and 111K tok/s end-to-end (Source: @gabriberton).
