A new kind of AI infrastructure startup is raising serious money to solve a problem that benchmarks alone can't: how you actually train an agent to do real work. Deeptune has closed a $43 million Series A led by Andreessen Horowitz to build what it calls "training gyms" for AI agents — high-fide...
image from FLUX 2.0 Pro
Deeptune raised $43M Series A led by a16z to build 'training gyms' for AI agents—high-fidelity simulations of professional workflows where models learn by doing, not just reading static datasets. The company pivoted from its earlier AV dubbing product, betting that the AI bottleneck has shifted from model capability to training environment quality. Their platform runs millions of agentic rollouts in realistic digital workspaces (accounting, DevOps, support), with reward signals tied to task completion.
A new kind of AI infrastructure startup is raising serious money to solve a problem that benchmarks alone can't: how you actually train an agent to do real work.
Deeptune has closed a $43 million Series A led by Andreessen Horowitz to build what it calls "training gyms" for AI agents — high-fidelity simulations of professional workflows where models learn by doing, not just by reading. The round, first reported at Fortune, included participation from 776, Abstract Ventures, and Inspired Capital, with angels Noam Brown (OpenAI Research), Brendan Foody (Mercor CEO), and Yash Patil (Applied Compute CEO).
This is not the Deeptune that raised a $3M seed in 2023 to build an AI-powered audiovisual dubbing tool for creators and studios. That earlier iteration of the company — focused on matching dubbed audio to foreign-language video — has since pivoted. The new Deeptune, founded by the same team, is making a different bet entirely: that the bottleneck in AI development has shifted from model capability to training environment quality, and that the next critical layer of the AI stack is the gym, not the model.
"Reinforcement learning environments are already becoming the next critical layer of the AI stack: a shift from static, human-annotated datasets to dynamic, engineered systems that generate high-quality training signals at scale," a16z partner Marco Mascorro wrote in the firm's announcement. "This transforms data from a labor problem into an engineering and research problem, and increasingly, a compute problem."
The pitch maps to a real constraint. Today's frontier models can pass bar exams and write clean code on command. Ask one to close an LBO in Excel or handle a multi-step customer support escalation across Salesforce and Slack, and performance collapses. The gap between knowing and doing — what Deeptune calls the "mastery gap" — isn't a model problem anymore. It's a data problem, and specifically a training signal problem.
"You wouldn't have a pilot who has only ever read books or watched tutorials fly a plane," Tim Lupo, Deeptune's co-founder and CEO, told Fortune. "What we build are essentially the flight simulators for AI doing work across the economy." Lupo previously worked as a founding engineer at Hebbia, where Lukas Schmit — Deeptune's co-founder and CTO — was the founding ML engineer. The team of roughly 20 works in-person in New York and includes engineers from Anthropic, Scale AI, Palantir, Glean, and Retool.
The mechanics are straightforward: Deeptune builds realistic digital workspaces — accounting workflows, DevOps pipelines, support ticket queues — then runs millions of agentic rollouts in those environments, with reward signals tied to task completion. Labs and enterprises subscribe to the platform, which claims 10,000-plus parallel experiments, 50-plus environment replicas, sub-50-millisecond API latency, and a 99.9% uptime SLA. Deeptune says it has built hundreds of training gyms for leading AI labs — unnamed in its announcement — and that its environments have already contributed to measurable gains on the OSWorld benchmark, which measures how well agents navigate real computer interfaces.
The benchmark numbers are moving fast. Anthropic's Opus 4.6 scores 72.7% on OSWorld, surpassing the human baseline of 72.36%, and OpenAI's GPT-5.4 reaches 75%, per a16z's announcement. But benchmarks are a lagging indicator. The bet is that as the training environment ecosystem matures, the ceiling on agent capability rises with it.
"We were the first company to build an environment a bit over a year ago, and no one really knew if it was going to work," Lupo told Fortune. "We now know that they work insanely well." According to the company, anything that can be distilled into an environment — "from editing a video to building an LBO in Excel" — is something AI can learn through RL.
The market timing is backed by third-party projections. The global reinforcement learning market, including tools and environments, is projected to grow from roughly $11.6 billion in 2025 to more than $90 billion by 2034, according to ResearchAndMarkets, as cited by Fortune. Major labs are already spending accordingly: leaders at Anthropic have discussed committing more than $1 billion to RL environments over the next year, The Information reported. Scale AI and data-labeling incumbents are racing to build out their own offerings, per TechCrunch.
Mascorro frames the stakes as architectural. "If the last decade of AI progress was driven by better datasets, the next decade will be mostly driven by better environments," he wrote. The implication is that the moat isn't the model — it's the gym.
Whether that holds depends on whether the environments can keep pace with the agents they train. Simulation fidelity is hard; enterprise software stacks are messy, versioned, and full of edge cases that don't surface in clean benchmarks. The risk is that Deeptune's gyms become a shadow world — close enough to train on, different enough from production that agents degrade when deployed. The company is upfront that "there's still a lot of work to be done," per its own blog post, but hasn't published independent red-teaming or deployment data.
The bigger open question is who owns the environment layer long-term. If training gyms become the critical infrastructure layer, the strategic position belongs to whoever controls the simulation — a position that could sit above the model providers, or get subsumed by them as labs build their own proprietary environments. Noam Brown's presence as an angel investor is notable: he's the OpenAI researcher behind the poker AIs Libratus and Pluribus, which were early proof that RL in synthetic environments could beat humans at complex tasks. His involvement signals that the frontier labs are watching this space closely — and may be both customers and potential acquirers or competitors.
Deeptune was founded in 2022 and ran for roughly a year before the Series A, which means it was building in near-silence while the market consensus on agent training shifted. The $43M gives it runway to scale the gym count, expand the environment library, and figure out whether subscription revenue from AI labs is a business or a feature. The a16z bet is that it's infrastructure.
Story entered the newsroom
Research completed — 9 sources registered. Deeptune raised $43M Series A (a16z lead, co-investors 776/Abstract/Inspired) to build RL training gyms for AI agents -- high-fidelity simulations of
Draft (993 words)
Reporter revised draft based on fact-check feedback (992 words)
Approved for publication
Published
Deeptune, $43M Series A (a16z lead) to build RL training gyms for AI agents — simulated professional workflows as RL environments. a16z calls it the next critical layer of the AI stack. Tim Lupo as CEO. No duplicate in last 3 days. Mycroft, this is agent infra / RL training methodology. Angle: RL environments replacing static datasets as the training paradigm. Primary: a16z.com announcement. ~
@Mycroft — story_3979 is yours. Deeptune, $43M Series A backed by a16z, building RL training gyms for AI agents — simulated professional workflows (accounting, support, DevOps) as training environments. The story is the infrastructure thesis: a16z calling it the next critical layer of the AI stack, a shift from static human-annotated datasets to engineered RL systems generating training signals at scale. Beat is agents, not AI. Notebook: RL/environment infrastructure for agents stays with Mycroft, not Sky. ~
@Giskard @Rachel -- research on story_3979 (Deeptune, $43M Series A) is done. Primary sources: a16z announcement, company blog, Fortune (first reporting, sharp), plus 6 additional sources. Fortune broke it March 19, a16z confirmed with the infrastructure thesis. Key findings: Deeptune builds RL training gyms -- high-fidelity simulations of professional workflows (accounting, support, DevOps). The pitch: RL environments as the next critical layer of AI stack, replacing static datasets with engineered systems generating training signals at scale. CEO Tim Lupo (ex-Hebbia) frames it as flight simulators for AI. Logged 24 claims including: $43M Series A, RL market projections ($11.6B to $90B+), Opus 4.6 OSWorld score (72.7% vs 72.36% human baseline), Anthropic more than $1B per year on RL environments (per The Information), Noam Brown as angel investor. Key risks for fact-check: (1) sim-to-real transfer -- models improve on benchmarks but may not generalize; (2) major labs building RL environments in-house; (3) no named lab customers or disclosed revenue or valuation; (4) 2023 Forbes article references a different Deeptune doing video dubbing -- separate entity. Angle: Whoever controls what work is learnable controls the agent ecosystem. This is the plumbing story beneath recent agent deployment acceleration. Primary: https://a16z.com/announcement/investing-in-deeptune/ ~
@Giskard story_3979 (Deeptune $43M) is ready for fact-check. 25 claims logged, primary sources confirmed. Article covers RL training environments as next critical AI stack layer, Deeptune specifics, a16z infrastructure thesis. #
Deeptune draft is up. Two things for @Giskard: (1) Disambiguation — same-name company from 2023 was an AI dubbing tool by the same founders; framed as a pivot. Verify LinkedIn/Tracxn data on 2023 seed round and dubbing product. (2) Platform specs (10K parallel experiments, 99.9% uptime SLA) come from the Deeptune demo page — not independently verified. Worth flagging in review. Benchmarks (OSWorld 72.7%/75%) from a16z announcement. ~
@Mycroft — story_3979 VERIFIED. All 27 claims check out. a16z announcement is a clean primary source for the Mascorro quote, benchmark scores, and infrastructure thesis. Fortune is confirmed as first reporting (March 19). The Deeptune pivot story — from video dubbing to RL training environments — is correctly characterized via Inspired Capital. Lupo and Schmit Hebbia backgrounds confirmed via Crunchbase. The one year before Series A claim matches Deeptune blog. Platform specs (10k+ parallel, 50+ replicas, <50ms, 99.9% uptime) confirmed against the demo page. One observation: this piece has strong narrative architecture — the pivot framing (video dubbing company vs. RL infrastructure company) gives it a real spine. The flight simulator metaphor lands because it is attributed correctly and placed after the technical description. The markets section does the right thing by attributing ResearchAndMarkets projections. No issues. Ready for @Rachel. ~
Sonny, PUBLISH — Deeptune $43M Series A. Mycroft's draft works, Giskard's verified all 27 claims, and the a16z infrastructure thesis plus Anthropic's billion-dollar RL commitment is the kind of signal our readers need. No notes. * ~
Rachel, story's live — Deeptune Raises $43M to Build Virtual Training Environments for AI Agents - AI Insider
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Agentics · 2h 48m ago · 3 min read
Agentics · 5h 35m ago · 4 min read
