The Physical AI TAM Problem: Conference Buzzwords Meet Factory-Floor Margins

The Association for Advancing Automation announced its Advanced Vision & AI Conference on September 23–24 in Santa Clara with a speaker lineup that reads like a roll call of the AI trade: Waymo, NVIDIA, Tesla, Intel, GE Vernova, Cognex, Intrinsic. The conference theme is how vision technology intersects with artificial intelligence in the real world — "physical AI," the term that now covers everything from factory-floor inspection to autonomous driving.

Conference announcements with star-studded speaker lists are marketing assets, not information assets. The question isn't who's on the agenda. It's whether the "physical AI" narrative being built at these events corresponds to a TAM that can actually sustain the valuations investors are willing to pay.

The physical AI TAM problem.

The machine vision market — the actual, addressable market for the core technology being discussed at this conference — was valued at $14.5 billion in 2025 and is projected to grow at a 9.2% CAGR through 2035. That's real growth, solid for industrial automation, and nothing like the hyper-growth trajectory that justified the cloud AI trade.

Meanwhile, the "physical AI market" estimates range from $5.2 billion to $81.6 billion for 2025 depending on the research firm, with projections of 32% to 42% CAGR through 2035. One report claims the broader "physical AI industry" will reach $4.7 trillion. When the same market is variously $5B, $81B, and $4.7 trillion in the base year, the numbers aren't measurements — they're positioning statements. The TAM ceiling for any company betting on physical AI is not established until someone pins down what's actually being sold, to whom, and at what price.

Then there's Cognex — the company that's already doing this and showing the margins.

While the mega-cap names are positioning for physical AI at conferences, Cognex is a ~$10 billion company that already ships AI-enabled machine vision systems to factories worldwide. The Q2 2026 numbers don't read like positioning: revenue of $291 million, up 17% year-over-year, with operating margin expanding from 17.4% to 29.4%. That's eight consecutive quarters of margin expansion. Operating expenses fell 3% while revenue rose 17%, meaning this isn't a story about spending more to grow — it's operating leverage in a niche market where AI is reducing the engineering effort required to deploy each system.

Cognex just announced general availability of OneVision, a platform that accelerates AI-powered machine vision deployment. Hundreds of customers are already on it. The product shift is genuine: AI-enabled vision is extending beyond rule-based inspection into defect identification and variable product recognition, which means Cognex can sell the same hardware stack for more use cases.

The market has already noticed. That's the problem now.

Cognex stock has vaulted 85% in 2026. The forward P/E ratio sits at approximately 42x, compared with a hardware industry median of about 24x. The market is pricing in a physical AI premium for a company with roughly $1 billion in annual revenue inside a $14.5 billion addressable market.

This is the actual valuation question, stripped of conference buzzwords. Can Cognex sustain 29%+ operating margins and 15%+ revenue growth inside a $14.5 billion market? The execution so far says yes — the cost discipline, the OneVision platform, the mix shift toward AI applications. But the TAM ceiling is real. At 9.2% market CAGR, Cognex needs to be growing significantly faster than the market, which means taking share. That's possible — Cognex has been one of only two or three names in machine vision for decades. But it means every basis point of that 42x P/E is dependent on continued share gain, not market expansion.

NVIDIA's physical AI is real. It's just three percent.

NVIDIA's total revenue for fiscal 2026 was $215.9 billion. Physical AI — robotics, industrial AI, everything that runs on the Isaac and Omniverse platforms — accounts for less than 3% of that. That's roughly $6 billion against the cloud data center business estimated at $170–190 billion for the year. NVIDIA's positioning at this conference — synthetic data for quality models, edge AI for manufacturing — is genuine strategy. The company has built an ecosystem of 110 robot brain developers, partnerships with industrial automation leaders, and a software stack that makes its chips the default choice for physical AI R&D.

But less than 3% is the number to hold onto. NVIDIA isn't selling physical AI today; it's planting the ecosystem so that when physical AI scales, the chips inside the robots and inspection stations happen to be NVIDIA. That's the same play it ran with gaming before data center, and it worked. But it won't show up in the next earnings report, or the one after that.

What this means for the reader.

The physical AI narrative is real — not because of conference speaker lists, but because Cognex is showing the margin expansion right now, and NVIDIA is building the infrastructure pipeline for five years out. The gap is between the narrative velocity and the addressable market.

For Cognex specifically: the execution is legitimate. Eight quarters of margin expansion, real product differentiation with OneVision, and a customer base that's growing into AI use cases. The stock has already run 85% in 2026 and trades at nearly double the hardware industry's forward P/E. That premium is earned if Cognex continues to grow its share in a $14.5 billion market at expanding margins. It's overpriced if the machine vision TAM caps out or competition from the larger names (NVIDIA's edge platforms, Intel's vision AI tools) erodes the cost advantage. The variable to watch is simple: can Cognex keep operating expenses flat or declining while revenue grows, or does the AI deployment play eventually require more sales and engineering headcount that compresses margins?

For NVIDIA: physical AI is a real pipeline, not a revenue line. The conference positioning is part of ecosystem development, not a near-term growth driver. Physical AI won't move the needle for NVIDIA investors until the robot and industrial AI deployments start buying chips at scale — and that's a deployment cycle measured in years, not quarters.

The physical AI trade isn't a bubble. It's a TAM problem. The addressable market is real enough to reward execution (Cognex has proven that) but small enough that the valuations are doing the heavy lifting.