CNBC Technology

Three AI Leaders on the New AI Fight: Controlling Cost, Agents, and Robots

Jeetu Patel, Lin Qiao & Evan Beard· Cisco President; Fireworks AI CEO; Standard Bots CEO at Cisco, Fireworks AI & Standard Bots
·~47 min·English·CNBC
AgentsInferenceRoboticsBusiness StrategyAI Infrastructure
TL;DR

Three CNBC interviews across the stack — Cisco, Fireworks AI, and Standard Bots — show the AI fight shifting from raw capability to control over agents, cost, and the physical world.

01The Frame

The Next Fight Is About Control, Not Capability

The week's biggest AI story wasn't a new model but the market forcing a question about control over what AI does, what it costs, and what it reshapes.

AI is moving from answering questions to taking action, and as it spreads, control is becoming the next big fight.

— CNBC Technology
Key Insight
The segment reframes the AI debate: until now the dominant question was how capable models could get. Now that agents act, money tightens, and robots ship, the binding constraint is governance — who controls the behavior, the bill, and the platform.
02Chip Economics

The Compute Center of Gravity Moved to Inference

Most compute now serves inference, not training, which extends the earning life of older GPUs instead of retiring them.

60% of all compute volume now has gone to inferencing

— Jeetu Patel, CNBC Technology
Key Insight
This inverts the bearish 'GPUs age like iPhones' math. If inference keeps growing and older chips absorb that load, a data center's hardware can get older while its earning power rises — but it also concentrates the best chips with the labs that extract the most value per FLOP.
03Agent Safety

Agents Are Teenagers With No Fear of Consequences

As agents act autonomously, a security breach and an agent simply misbehaving become nearly impossible to tell apart, so security and observability have to fuse.

The way I think about this is, agents are like teenagers. They're supremely intelligent.

— Jeetu Patel, CNBC Technology
Key Insight
Patel's real claim is operational, not cute: the old security stack assumes a human attacker. When the 'attacker' might be your own agent following its goal too literally, the fix isn't a firewall — it's runtime observability that detects drift and enforces guardrails while the agent is still running.
04Cost Control

A $50,000 Bill for a $20 Task

An unsupervised agent can spawn thousands of sub-agents chasing every route, so cost overrun becomes a runtime safety problem, not just a billing surprise.

You could get a bill for $50,000 when you thought it was going to be a $20 task

— Jeetu Patel, CNBC Technology
Key Insight
Cost stops being a finance-team problem and becomes an engineering one. If an agent can recursively spend money at machine speed, the budget has to be enforced in the loop — the same runtime guardrail that stops bad behavior also has to stop bad economics.
05Unit Economics

The AI Era Invented a New Way to Go Broke

Unlike SaaS, where scaling cheap software meant profit, scaling an AI product means scaling GPU spend, so more usage can burn more cash.

there's really a phase we are seeing is scaling into bankruptcy. This is such a unique phenomenon in the AI era that is fundamentally different from traditional SaaS era

— Lin Qiao, CNBC Technology
Key Insight
This is why 'blitz-scaling' quietly died in AI. When each new active user adds GPU cost that never falls to zero, growth without margin discipline is a countdown, and it's worst for large incumbents who can't roll AI out to their whole base without torching cash flow.
06Strategy

From Renting Tokens to Owning Your Intelligence

Qiao's bet is that companies stop measuring success by tokens consumed and start turning their private data into specialized models they own.

we see a big shift of industry from token maxing to value maxing

— Lin Qiao, CNBC Technology
Key Insight
The strategic tell is 'own your intelligence.' If a company's edge lives in proprietary data, renting a general frontier model per token both leaks margin and leaves that edge un-activated — so the moat becomes a specialized model distilled from data no competitor has.
07Robotics

America Deploys One Robot for Every Ten China Does

China deployed roughly ten times as many robots as the US last year and controls the component supply chain, leaving American factories far behind on cost and scale.

the United States deployed 1/10 the number of robots that China did last year. China deployed about 300,000. We were closer to 30,000.

— Evan Beard, CNBC Technology
Key Insight
Beard's framing echoes the open-source debate one segment earlier: banning the cheapest, most capable imports can protect a nascent domestic industry or handicap the companies that need those robots now. The bet is that a fair playing field plus American manufacturing beats raw cost — the same wager the US is making on chips.
08The Platform Bet

Robotics Is 1970 for the Computer Industry

Beard argues robotics today looks like personal computing in 1970 — obviously useful, barely deployed, and about to become the next platform whoever owns it will lead.

robotics is like 1970 for the computer industry. It's very nascent.

— Evan Beard, CNBC Technology
Key Insight
This is where all three interviews rhyme. Control over agents, over cost, and over robots is really one argument about owning platforms early — because whoever controls the next platform, the way the US once controlled computing, sets the terms for everyone who builds on it.