Jensen Huang Just Told Every Company What to Build. Most Aren’t Listening.

THE NUMBER: 250,000 — GitHub stars for OpenClaw in weeks, not years. Jensen Huang called it the most successful open-source project in history and the operating system for personal AI. Every enterprise company, he said, needs an OpenClaw strategy. But the real question isn’t whether you have one. It’s whether your business can even be read by one.

At GTC last week, Jensen Huang didn’t just announce products. He announced a new competitive requirement. Every company needs a claw strategy — a plan for deploying AI agents and, just as critically, a plan for making their business accessible to the agents everyone else is deploying.

That second part is what nobody’s talking about.

OpenClaw went from a weekend project to a quarter-million GitHub stars and an NVIDIA enterprise platform (NemoClaw) in weeks. It’s now part of OpenAI, which may create trust issues for enterprises that were betting on its neutrality. NemoClaw, built with NVIDIA’s Nemotron models and OpenShell runtime, will likely be better received — it adds privacy and security controls that enterprise buyers actually care about.

But the infrastructure announcement isn’t the story. The forcing function is. Because when Jensen says every company needs a claw strategy, what he’s really saying is: there will soon be far more agentic users of your services than human ones. And agents don’t care about your beautiful UI. They have no eyes, no emotions, no patience for your seventeen-step checkout flow. They need one thing: clean, structured access to your data. Can I get what I need or not?

Meanwhile, Tomasz Tunguz published the sharpest analysis of the week: AI agents are now commanding 75-100% of human equivalent salary in labor-shortage markets. Goldman Sachs found that low-labor-cost stocks outperformed by eight percentage points. Labor’s share of GDP just hit a record low of 53.8%. And here’s the part Tunguz undersells: agents don’t pay FICA, don’t need benefits, don’t file for unemployment. That’s a 25-30% cost reduction before you even count the productivity gain.

And Sam Altman made the subtext into text, telling audiences that AI will be sold the way utilities sell electricity — metered, usage-based, tokens on a bill. He’s probably right about enterprise. He may have also just handed the best pitch deck the open-source and on-device AI movement has ever received.

Your Agent Can’t Read Your Website. That’s a Bigger Problem Than You Think.

There’s a phrase making the rounds in enterprise circles right now: “agent-readable and agent-writable.” It sounds like an API problem. It’s actually a data architecture problem that runs down your entire stack.

Think about how a customer buys from you today. They browse your site, compare options, read reviews, maybe call someone. A human can navigate ambiguity. They can figure out that “Premium” and “Enterprise” are actually the same tier with different names. They can forgive a broken search and try a different query.

An agent can’t. An agent needs structured, consistent, machine-readable access to every product attribute, every price point, every availability status. And here’s what Nate’s Substack nailed this week: 80% of what makes your product worth buying lives in people’s heads, not your databases. The tribal knowledge — “oh, the medium runs large” or “this SKU is actually discontinued but we keep it listed” — is invisible to an agentic buyer.

The companies that figure this out first don’t just serve agents better. They build a structural advantage that compounds every quarter. Every clean data path you create is a moat against competitors whose systems are still optimized for eyeballs. The irony is beautiful: the path from outside the Fortune 500 to inside it may not run through McKinsey. It may run through making your business legible to machines before your competitors do.

And let’s be honest about who this affects most. If you’re a $20M revenue company, McKinsey isn’t coming. PwC — which just told its staff there’s “no opportunity to opt out” of AI — is too expensive and too busy serving the Fortune 500. You’re going to have to figure out your claw strategy yourself. That’s not a disadvantage. Smaller companies move faster. The question is whether you move fast enough.

The action item: This week, try to buy your own product using an AI agent. Watch where it breaks. That’s your roadmap.

Agents Don’t Need a Salary. They Need a Tax Code.

Tunguz’s numbers are stark, but they only tell part of the story. An agent at 75% of a human salary isn’t just cheaper. It’s structurally different.

When you hire a person, you’re paying for about 1,800 productive hours per year after vacation, sick days, meetings, and the slow Wednesday after the all-hands. An agent works 8,760 hours. It doesn’t need to be managed in the traditional sense — no performance reviews, no career development conversations, no retention risk that keeps you up at night.

But here’s where the analysis gets interesting. The conventional framing is: you have 100 people, AI lets you cut 50, margins go up. And sure, that math works. But it’s the wrong math.

What if you keep 25 of those 50 — or hire a different 25 with different skills — and give each of them 10 agents? Or 20? Now you haven’t just cut costs. You’ve multiplied capacity by an order of magnitude. You can cover territory you never could before. Enter markets that were uneconomical. Serve customer segments that didn’t justify the headcount.

The efficiency equation isn’t linear. Human-to-human coordination is relatively inefficient — meetings, emails, misunderstandings, context-switching. Agent-to-human is better but still clunky (anyone who’s watched an AI agent try to navigate a website through Chrome can attest to that). But agent-to-agent, with proper data infrastructure? That’s where the real unlock lives. And nobody’s debating that transition enough.

The companies that will win the next five years won’t be the ones that replaced the most humans. They’ll be the ones that figured out the right ratio of humans to agents and built the infrastructure for agents to talk to each other without friction.

Why it matters: Stop thinking about AI as a headcount reduction tool. Start thinking about it as a force multiplier. The question isn’t “how many people can I replace?” It’s “what could my best 25 people accomplish with 200 agents behind them?”

The Inference Subsidy Is Ending. Cursor Already Got the Memo.

Right now, every major frontier model is being sold below cost. OpenAI, Anthropic, Google — they’re all subsidizing inference to win market share, the same way Uber subsidized rides to kill taxis. That’s not a business model. It’s a land grab with an expiration date.

Sam Altman said the quiet part out loud: intelligence will be metered. Tokens on a bill. And the moment you tell customers the meter is coming, you hand the business case to everyone building an alternative to your meter.

Cursor read the room before anyone else. They shipped Composer 2 this week — their own frontier-competitive coding model that scores 61.7% on Terminal-Bench 2.0, beating Claude Opus 4.6 at 58%, at roughly one-tenth the cost. An application-layer company built a model that outperforms the lab that was supposed to be unbeatable at coding tasks.

This isn’t an anomaly. It’s a preview. Cursor had to make this move. Claude Code and OpenAI’s Codex were growing too fast, threatening to commoditize the very layer Cursor monetizes. But the strategic lesson applies everywhere: when your supplier tells you they plan to charge you more, the rational response is to build your own supply.

This is exactly what happened with solar panels. The moment electric companies signaled rate increases, a meaningful chunk of customers started pricing out alternatives. Not everyone. But enough to matter. NVIDIA’s DGX Spark, Mac Studios with enough unified memory to run capable models locally, open-source models that keep closing the gap — these are the solar panels of the AI era. Think Alexa won’t have real AI built into the silicon soon? Siri?

The frontier labs have about 18 months before the application layer and the device manufacturers start eating their margin from both sides.

The signal: If you’re building on a single frontier model’s API today, you’re building on a subsidy. Start prototyping with open-source alternatives now — not as your primary stack, but as your insurance policy.

What This Means For You

Jensen’s GTC wasn’t a product launch. It was a declaration: the agentic era is here, the infrastructure layer is set, and the companies that aren’t building for it are already behind. The question every allocator of human and physical capital should be asking isn’t “should we use AI?” It’s “can AI use us?”

Make your business agent-readable before your competitors do. The company whose data infrastructure lets agents transact cleanly will take market share from the company that’s still optimized for human eyeballs. This isn’t a 2028 problem. Agents are transacting now.

Rethink headcount as a ratio, not a number. The crude “replace humans with AI” math is wrong. Twenty-five people with two hundred agents behind them will outperform a hundred people with no agents every single time. The right question is: what’s your human-to-agent ratio, and what infrastructure do those agents need to work together?

Diversify your inference dependency. The subsidy era is ending. Cursor built its own model. Your company probably can’t. But you can architect your systems so you’re not locked into a single provider when the meter turns on. Multi-model, on-device capability, open-source fallbacks — build the optionality now while it’s cheap.

The companies that figure out their claw strategy first won’t just survive the agentic era. They’ll define it. And Jensen? He doesn’t care which strategy you pick. He’s selling the shovels to all of them.

Three Questions We Think You Should Be Asking Yourself

Could an AI agent buy your product right now without a human in the loop? Try it. Seriously — this afternoon. Use Claude or ChatGPT to attempt a purchase or engagement with your business. Every friction point you hit is a friction point that will cost you agentic customers within 12 months. The businesses that are legible to machines will eat the ones that aren’t.

What would your company look like with half the people and ten times the agents? Not a layoff fantasy — a genuine architectural question. If your best 25 employees each had 10-20 AI agents working for them, what markets could you enter that you can’t today? What customer segments become economical? The answer to that question should be driving your hiring, training, and infrastructure decisions right now.

What happens to your unit economics when inference isn’t subsidized? Every API call you’re making to Claude or GPT today is priced below cost. When the meter turns on — and Altman just told you it will — does your business still work? Run the math at 3x current token prices. If the answer is uncomfortable, you need a fallback plan before you need it.

Every single enterprise company and every single software company in the world needs an agentic strategy, and specifically needs an OpenClaw strategy.”

— Jensen Huang, NVIDIA GTC 2026

— Harry and Anthony

Sources:

• Shelly Palmer: Corporate Claws
• Nate’s Substack: Your Systems Are Not Agent-Ready
• Mindstream: Altman May Have Made a Big Mistake
• The Deep View: The Race for Power Behind AI’s Utopian Story
• NVIDIA NemoClaw Announcement
• Dylan Patel on the AI Chip Race
• Tomasz Tunguz: The Pricing Power of Agents