Poetic hit eight-figure revenue with 4 people. Now Kleiner Perkins is betting it's the future of enterprise AI.

A $3.5B War Chest, and What It Signals

Kleiner Perkins closed $3.5 billion across two new funds on Tuesday, the firm's largest single fundraise and its most concentrated AI deployment to date, according to TechCrunch. The haul splits into $1 billion for KP22, its 22nd early-stage vehicle, and $2.5 billion targeting late-stage growth companies, a 75% increase over the $2 billion the firm raised less than two years ago.

The structure is deliberate. Early-stage checks seed the next cohort of AI-native startups; the growth tranche lets Kleiner Perkins double down on winners that have moved past product-market fit and are spending heavily on compute, headcount, and deployment. That two-tier setup is built to capture AI companies from first check through IPO, and it gives the firm enough dry powder to compete with the mega-funds circling the sector. Thrive Capital raised $10 billion in fresh commitments this February. General Catalyst is targeting a similar figure. Founders Fund closed $6 billion for its fourth growth vehicle, an SEC filing confirmed.

Against that backdrop, Kleiner Perkins' new capital is not the largest number in the market. But it is the most AI-dedicated pool of that size among the legacy firms that have been in the game long enough to have real pattern recognition. The firm backed Google, Amazon, and Uber before those names meant anything. Now it is making the same concentrated bet on AI — across foundation models, infrastructure, and deployment — with the explicit thesis that the winners will need venture partners who can operate at every stage.

The timing is not accidental. Kleiner Perkins has recent exits to point to. Figma's IPO last year returned significant capital on a $25 million Series B the firm led in 2018. Google's acqui-hire of portfolio company Windsurf added another liquidity event. Those wins give the new fund a track record to show LPs, and they give the investment team (now just five partners, after Ev Randle's departure to Benchmark and Annie Case's move to an advisory role) the credibility to be selective rather than spray capital across a sector that is attracting more money than disciplined founders.

For the AI engineers and operators weighing their next move, the fundraise is a signal of where the hiring demand is heading. When a firm raises this much with an AI-only thesis, the portfolio companies will need to staff up. The jobs are coming. The question is who gets there first.

The Apple Engineer's $80M Infrastructure Bet

Sail Research, a San Francisco-based startup building infrastructure for autonomous AI agents, has emerged from stealth with $80 million in combined seed and Series A funding at a $450 million valuation, Fortune reported. Kleiner Perkins led the Series A. Sequoia led the seed round. Redpoint Ventures, Theory Ventures, Vine Ventures, CRV, A*, and Abstract Ventures also participated, along with individual checks from Stanford's John Hennessy, Cadence's Lip-Bu Tan, and independent director Tri Dao.

The co-founder is Neil Movva, a 28-year-old former Apple engineer who worked on the chip powering computer vision across a billion iPhones before growing frustrated that Apple's AI ambition, in his view, capped out at animoji. His co-founder and CTO, Samir Menon, also came from Apple's security engineering organization. The two met on the first day of freshman year at Stanford, took the same course load, and reunited in late 2025 to rebuild the inference stack from scratch.

Movva's thesis is blunt: AI infrastructure was designed for the wrong problem. The current generation of inference platforms optimizes for low latency, which works fine for a chatbot that answers a question in two seconds. But enterprises are increasingly deploying agents that run autonomously for hours, reading entire codebases, screening hundreds of job candidates, or conducting research without a human in the loop. Agentic workflows consume tokens at a rate 50 to 500 times higher than simple chat. Per-token prices have fallen, yet enterprise AI bills have tripled. Goldman Sachs forecasts a 24-fold increase in token consumption by 2030.

Sail's answer is to optimize for throughput instead of latency, packing far more compute work into every unit of power by writing scheduling and optimization software that sits directly on top of existing GPUs. Movva claims customers typically see 3x to 10x cost improvements over comparable alternatives. The tradeoff is deliberate, he told Fortune: "We only care about efficiency. It's quite difficult to build an inference engine for both throughput and latency at the same time. Everyone else is optimizing for latency, and we just care about throughput."

One early customer, Detail.dev, uses Sail to run code-review agents that spend three to four hours, sometimes longer, scanning entire codebases for bugs that five-minute reviews miss. Sail launched its inference service in March and already processes trillions of tokens per week.

The competitive picture is complicated. Together AI, a Kleiner Perkins portfolio company and one of the leading open-source inference providers, is a direct incumbent. Kleiner Perkins partner Aditya Naganath, who led the Sail investment, argued the two companies occupy different markets: Together owns interactive, chat-based workloads. Sail owns the long-running agent workload. The bigger threat may come from Anthropic, OpenAI, and Google, which are building their own inference infrastructure and could commoditize the layer.

Movva's counter: token prices have been flat or rising for six months, compute demand is growing faster than supply, and someone needs to obsess over squeezing the most intelligence out of every available GPU. "We feel an emotional pain when we see a GPU be idle or wasted in any way," he said.

For AI engineers weighing their next move, the signal is clear. Kleiner Perkins is betting that the inference layer, purpose-built for autonomous agents, is a market that could be 10x to 100x bigger than it is today. Sail's round is a concrete data point that deep infrastructure talent (the kind that understands both chip-level constraints and agent-level workloads) is commanding early-stage premiums that would have been unthinkable two years ago.

The Google/Waymo Exodus

Markie Wagner spent years watching AI systems break the moment they touched a real production workflow. She ran that pattern into the ground at Google and Waymo, where she worked as a machine learning engineer, and decided the gap between a working demo and a working product was wide enough to build a company in. Kleiner Perkins, OpenAI, and Founders Fund agreed, leading Poetic's $50M Series A at a $500M valuation in June 2026.

The round is the clearest signal yet that production-grade AI deployment, not model research or chatbot wrappers, has become its own venture category. Poetic's pitch is blunt: most enterprise AI pilots fail. MIT's NANDA initiative found that 95% of enterprise generative AI pilots never produce measurable financial results. Wagner built Poetic to go after the processes that banks and insurers actually run, rule-heavy and zero-tolerance-for-error work that generic AI agents handle badly.

The early numbers back her up. At SoFi, Poetic's platform runs fraud investigations end-to-end at 99% or higher accuracy, and reached production in five weeks. At AIG, it hit the same accuracy threshold on a multi-hour insurance workflow. The company says every pilot it has run has converted to a full production deployment, and hit an eight-figure annual revenue run rate in 2025 with a team of four. Current clients include SoFi, AIG, and Chime.

Wagner's background is part of what convinced Kleiner Perkins to lead. A Thiel Fellow who previously ran an AI consultancy called Delphi Labs, she rebranded the company from Forge to Poetic and built a custom programming language for enterprise automation. The system lets operators describe a process in plain language, then translates it into structured, step-by-step instructions that behave more like code than open-ended AI. The practical result is consistency, the thing regulated industries need and standard large language model integrations rarely deliver.

The hiring implications are direct. Poetic plans to use the $50 million to grow its forward-deployed team, the engineers and specialists who work on-site with clients to build automations. Kleiner Perkins partner Leigh Marie Braswell said Poetic has recruited leaders from Palantir, UiPath, Ramp, Scale, and Retool. For AI engineers tired of watching demos die in production, that's the pitch: go build the layer that actually works.

Zero G Talent's board currently lists roles at Waymo, including senior ML engineers in simulation realism and fleet optimization, and OpenAI is actively hiring across model design, agent post-training, and safety operations. The talent pipeline runs both ways: engineers leave incumbents to fix the deployment gap, and the incumbents hire to close it.

Mirendil's $200M Seed and the a16z Crossover Signal

Mirendil, a San Francisco AI startup founded in early 2026 by former Anthropic researchers Behnam Neyshabur and Harsh Mehta, raised a $200 million seed round at a roughly $1 billion valuation. Andreessen Horowitz and Kleiner Perkins co-led the round, with NVIDIA participating as an investor. The founding team of 20 engineers and researchers includes people from Anthropic, xAI, Google DeepMind, and OpenAI.

The size alone forces a comparison. Traded: Venture Capital reported the median US seed round sits around $3-4 million. Mirendil raised roughly 50 times that before shipping a commercial product or publishing a benchmark. The company's stated mission (building AI systems that can autonomously run experiments, evaluate results, and iterate on research problems) places it in what investors call the "neo-lab" category: startups using frontier AI to drive domain-specific research in biology and materials science rather than building consumer applications.

The a16z–Kleiner Perkins co-lead matters more than the dollar figure. These two firms compete aggressively for frontier AI deals and rarely share a cap table at the seed stage. Their joint participation signals that the deal's risk profile (no revenue, no product, 20-person team) was acceptable to both, which tells you how much capital is chasing a narrow band of ex-frontier-lab talent. Kleiner Perkins' Mamoon Hamid described Mirendil as a frontier laboratory focused on advancing AI R&D. a16z's Matt Bornstein said the team is working on one of the major hyperscale challenges in AI, though neither firm disclosed technical specifics.

NVIDIA's involvement follows a pattern the chipmaker has repeated across 2025 and 2026: fund compute-intensive startups early to lock in future GPU demand for training and inference. Mirendil joins a growing list of ex-frontier-lab spinouts raising nine-figure seed rounds, including Safe Superintelligence (co-founded by Ilya Sutskever) and Thinking Machines Lab (led by Mira Murati).

For the talent market, the signal is direct: if you have a publication record at Anthropic or DeepMind and want to start something, two of the most competitive firms in Silicon Valley will co-lead your seed round and a hardware giant will back you before you've hired your twenty-first engineer.

Inside Poetic's $500M Valuation

$50 million on a $500 million valuation, for a company that hit an eight-figure revenue run rate last year with four employees. Kleiner Perkins led the Poetic Series A in June 2026, with OpenAI, Founders Fund, and First Harmonic participating. The round prices a deliberate architectural bet: that the future of production AI belongs to deterministic execution layered over stochastic models, not to autonomous agents reasoning freely through enterprise workflows.

Founder Markie Wagner, a Thiel Fellow and former ML engineer at Google and Waymo, built Poetic (formerly Forge) after watching AI systems fail the moment they met real operations. She left Waymo, founded the ML consultancy Delphi Labs, and kept watching the same failure pattern repeat. The problem wasn't the models. It was the architecture. Probabilistic output, no matter how impressive in a demo, could not satisfy the audit requirements of a regulated industry.

Software that learns like AI, runs like code

Poetic's technical answer is a proprietary programming language. Operators describe complex workflows in natural language; the platform compiles that logic into deterministic, near-tokenless execution. The system does the same thing every time given the same input. That constraint is the product. Autonomous agents drift. Prompts produce variable outputs. Poetic restricts the model's degrees of freedom on the way out, wrapping stochastic intelligence in reproducible code that produces an audit trail a regulator will accept.

The target is work that has defeated everything else: fraud investigations, transaction monitoring, compliance checks, insurance reviews. Multi-hour processes running thousands of times a day, carrying near-zero tolerance for error, governed by thousands of unwritten rules accumulated over decades.

SoFi CEO Anthony Noto confirmed that Poetic executed fraud processes end-to-end while improving quality metrics, giving members restored access immediately instead of after days. AIG saw similar accuracy on processes that previously required heavy manual effort.

Why the investor lineup matters

OpenAI rarely takes equity positions in startups that are not strategically tied to its own roadmap. Its participation signals that Poetic's deterministic, low-token execution model is complementary to the large language model ecosystem rather than competitive with it. Founders Fund's involvement closes a circle that started with Wagner's Thiel Fellowship. Kleiner Perkins partner Leigh Marie Braswell backed the company personally before it had a product, then led both the seed and the Series A, doubling down at a nine-figure valuation in a conviction statement rather than a standard follow-on.

"What Poetic has built is genuinely different — a platform that can execute the complex, high-stakes processes that large enterprises actually run, with accuracy that exceeds what human teams can deliver." — Leigh Marie Braswell, Partner, Kleiner Perkins

The full-stack AI engineer signal

Poetic's four-person, eight-figure run rate is extreme product leverage. But the company is now staffing up, specifically its forward-deployed engineering team. Braswell named the talent pedigree: leaders from those same five companies. That mix is the point. Poetic needs engineers who understand both model behavior and production systems, people who can work across the stochastic layer and the deterministic wrapper. The hardware-software convergence here is architectural: you are building a system that translates probabilistic intelligence into hardened, auditable execution at scale. That requires full-stack AI engineers who can operate across the entire pipeline, from model integration to compliance-grade deployment.

The enterprise AI market is full of pilots that never reached production. Poetic's 100% conversion rate, if it holds as the company scales beyond financial services into healthcare and government, is the clearest signal that the reliability problem has actually been solved. For engineers weighing offers between model labs and infrastructure startups, the question is whether you want to build the model or build the system that makes the model safe to deploy. Poetic is betting the talent market shifts to the latter.

What This Means for AI Engineering Talent

The hiring math has flipped. For three years, the gravitational pull in AI engineering ran one way: toward OpenAI, Google DeepMind, and Anthropic, where a top researcher could pull $20 million a year and the compute budget was effectively infinite. That pull is real and it has not disappeared. But Kleiner Perkins' portfolio build-out (45 deals in 2025 alone, concentrated across inference infrastructure, agentic workflows, and enterprise deployment) has created a second gravitational center, one where the work itself looks different.

The engineers Kleiner Perkins is backing are not training foundation models. They are building the infrastructure those models run on when they leave the lab. Neil Movva's Sail Research, staffed with ex-Apple chip engineers, is optimizing GPU throughput for long-horizon agents, the kind of work that requires understanding silicon-level constraints and inference scheduling simultaneously. Poetic's hardware-software convergence at a $500M valuation demands engineers who can think across the boundary between physical compute and model architecture. These are not roles that map onto a standard ML research job description.

"The belief that inference is going to be a 10x — even 100x — bigger market than it is today." — Aditya Naganath, Kleiner Perkins partner, to Fortune

That quote should reframe how an engineer evaluates the offer sitting in their inbox. At OpenAI, you might train the next model. At Google DeepMind, you might deploy it at scale inside Google's infrastructure. At a Kleiner Perkins-backed startup like Sail, Poetic, or Roadrunner, you build the system that lets someone else's model run efficiently across thousands of simultaneous agent tasks. The skills are specific: GPU kernel optimization, inference orchestration, agentic workflow infrastructure, hardware-software co-design. These are not the skills that get you a research paper. They are the skills that get a customer to see a 3x to 10x cost improvement on their compute bill.

The compensation gap is real but narrower than it looks. OpenAI's posted roles on Zero G Talent's board range from $266,000 to $445,000. Google DeepMind's packages for top researchers reportedly hit $20 million, though that figure concentrates at the senior research level. Kleiner Perkins-backed startups likely offer lower base compensation but meaningful equity in companies with aggressive valuation trajectories. Sail Research hit a $450M valuation at its first institutional round. For an engineer optimizing inference kernels, the right startup bet may pay more over five years than a DeepMind salary, with a risk profile that is high but calculated.

The practical takeaway: if your work is model training, alignment, or research, the incumbents still win on resources and prestige. If your work is the layer between a trained model and a production system that runs at scale without burning the customer's budget, Kleiner Perkins has just built you a cluster of employers that will pay a premium for exactly that skill set. Check the open roles at OpenAI, Waymo, and the Kleiner Perkins portfolio companies on that same board to see where the demand sits right now.

Working in AI? Zero G Talent tracks the openings: browse AI jobs, openings at OpenAI, Waymo and DeepMind, and the people building the field.