Earlier this week, 5 individuals who contact each layer of the AI provide chain sat down on the Milken International Convention in Beverly Hills, the place they talked with this editor about the whole lot from chip shortages to orbital knowledge facilities to the likelihood that the entire structure that undergirds the tech is flawed.
On stage with TechCrunch: Christophe Fouquet, CEO of ASML, the Dutch firm that holds a monopoly on the acute ultraviolet lithography machines with out which fashionable chips wouldn’t exist; Francis deSouza, COO of Google Cloud, who’s overseeing one of many greatest infrastructure bets in company historical past; Qasar Younis, co-founder and CEO of Utilized Instinct, a $15 billion bodily AI firm that began in simulation and has since moved into protection; Dimitry Shevelenko, the chief enterprise officer of Perplexity, the AI-native search-to-agents firm; and Eve Bodnia, a quantum physicist who left academia to problem the foundational structure a lot of the AI business takes with no consideration at her startup, Logical Intelligence. (Meta’s former chief AI scientist, Yan LeCun, signed on as founding chair of its technical analysis board earlier this 12 months.)
Right here’s what the 5 needed to say:
The bottlenecks are actual
The AI increase is operating into laborious bodily limits, and the constraints start additional down the stack than many could understand. Fouquet was the primary to say it, describing a “large acceleration of chips manufacturing,” whereas expressing his “sturdy perception” that regardless of all that effort, “for the subsequent two, three, perhaps 5 years, the market will likely be provide restricted,” which means the hyperscalers — Google, Microsoft, Amazon, Meta — aren’t going to get all of the chips they’re paying for, full cease.
DeSouza highlighted how massive — and how briskly rising — a problem that is, reminding the viewers that Google Cloud’s income crossed $20 billion final quarter, rising 63%, whereas its backlog — the dedicated however not but delivered income — practically doubled in a single quarter, from $250 billion to $460 billion. “The demand is actual,” he stated with spectacular calm.
For Younis, the constraint comes primarily from elsewhere. Utilized Instinct builds autonomy techniques for vehicles, vans, drones, mining gear and protection automobiles, and his bottleneck isn’t silicon — it’s the info that one can solely collect by sending machines into the actual world and watching what occurs. “It’s important to discover it from the actual world,” he stated, and no quantity of artificial simulation totally closes that hole. “There will likely be a very long time earlier than you’ll be able to totally practice fashions that run on the bodily world synthetically.”
Techcrunch occasion
San Francisco, CA
|
October 13-15, 2026
The vitality drawback can also be actual
If chips are the primary bottleneck, vitality is the one looming behind it. DeSouza confirmed that Google is exploring knowledge facilities in house as a severe response to vitality constraints. “You get entry to extra ample vitality,” he famous. After all, even in orbit, it isn’t easy. DeSouza noticed house is a vacuum, so eliminates convection, leaving radiation as the one technique to shed warmth into the encircling surroundings (a a lot slower and harder-to-engineer course of than the air and liquid cooling techniques that knowledge facilities depend on in the present day). However the firm remains to be treating it as a respectable path.
The deeper argument de Souza made, considerably unsurprisingly, was about effectivity by means of integration. Google’s technique of co-engineering its full AI stack — from customized TPU chips by means of to fashions and brokers — pays dividends in watts per flop that an organization shopping for off-the-shelf elements merely can’t replicate, he steered. “Operating Gemini on TPUs is way more vitality environment friendly than every other configuration,” as a result of chip designers know what’s coming within the mannequin earlier than it ships, he stated. In a world the place vitality availability is turning into an enormous constraint on how far this tech can go, that form of vertical integration is a serious aggressive benefit.
Fouquet’s echoed the purpose later within the dialogue. “Nothing could be priceless,” he stated. The business is in an unusual second proper now, investing extraordinary quantities of capital, pushed by strategic necessity. However extra compute means extra vitality, and extra vitality has a value.
A unique form of intelligence
Whereas the remainder of the business debates scale, structure, and inference effectivity throughout the massive language mannequin paradigm, Bodnia is constructing one thing very completely different.
Her firm, Logical Intelligence, is constructed on so-called energy-based fashions (EBMs), a category of AI that doesn’t predict the subsequent token in a sequence however as an alternative makes an attempt to know the foundations underlying knowledge, in a approach she argues is nearer to how the human mind really works. “Language is a consumer interface between my mind and yours,” she stated. “The reasoning itself shouldn’t be connected to any language.”
Her largest mannequin runs to 200 million parameters — in comparison with the tons of of billions in main LLMs — and she or he claims it runs hundreds of instances sooner. Extra importantly, it’s designed to replace its data as knowledge adjustments, somewhat than requiring retraining from scratch.
For chip design, robotics and different domains the place a system wants to know bodily guidelines somewhat than linguistic patterns, she argues EBMs are the extra pure match. “While you drive a automotive, you’re not looking for patterns in any language. You go searching you, perceive the foundations concerning the world round you, and decide.” It’s an fascinating argument and one which’s prone to entice extra consideration within the coming months, given the AI discipline is starting to ask whether or not scale alone is enough.
Brokers, guardrails, and belief
Shevelenko spent a lot of the dialog explaining how Perplexity has developed from a search product into one thing it now calls a “digital employee.” Perplexity Pc, its latest providing, is designed not as a software a data employee makes use of, however as a employees {that a} data employee directs. “Each day you get up and you’ve got 100 employees in your workforce,” he stated of the chance. “What are you going to do to benefit from it?”
It’s a compelling pitch; it additionally raises apparent questions on management, so I requested them. His reply was granularity. Enterprise directors can specify not simply which connectors and instruments an agent can entry, however whether or not these permissions are read-only or read-write — a distinction that issues enormously when brokers are appearing inside company techniques. When Comet, Perplexity’s computer-use agent, takes actions on a consumer’s behalf, it presents a plan and asks for approval first. Some customers discover the friction annoying, Shevelenko stated, however he stated heconsiders it important, significantly after becoming a member of the board of Lazard, the place stated he has discovered himself unexpectedly sympathetic to the conservative instincts of a CISO defending a 180-year-old model constructed fully on consumer belief. “Granularity is the bedrock of excellent safety hygiene,” he stated.
Sovereignty, not simply security
Younis supplied what could have been the panel’s most geopolitically charged remark, which is that bodily AI and nationwide sovereignty are entangled in ways in which purely digital AI by no means was.
The web initially unfold as American know-how and confronted pushback solely on the software layer — the Ubers and DoorDashes — when offline penalties turned seen. Bodily AI is completely different. Autonomous automobiles, protection drones, mining gear, agricultural machines — these manifest in the actual world in methods governments can’t ignore, elevating questions on security, knowledge assortment, and who finally controls techniques that function inside a nation’s borders. “Nearly persistently, each nation is saying: we don’t need this intelligence in a bodily kind in our borders, managed by one other nation.” Fewer nations, he advised the group, can presently discipline a robotaxi than possess nuclear weapons.
Fouquet framed it a bit of in a different way. China’s AI progress is actual — DeepSeek’s launch earlier this 12 months despatched one thing near a panic by means of components of the business — however that progress is constrained under the mannequin layer. With out entry to EUV lithography, Chinese language chipmakers can’t manufacture probably the most superior semiconductors, and fashions constructed on older {hardware} function at a compounding drawback irrespective of how good the software program will get. “At the moment, in america, you will have the info, you will have the computing entry, you will have the chips, you will have the expertise. China does an excellent job on the highest of the stack, however is missing some components under,” Fouquet stated.
The era query
Close to the top of our panel, somebody within the viewers requested the apparent uncomfortable query: is all of this going to impression the subsequent era’s capability for essential pondering?
The solutions had been, maybe unsurprisingly, optimistic, although not naively so. De Souza pointed to the size of issues that extra highly effective instruments would possibly lastly let humanity tackle. Assume neurological illnesses whose organic mechanisms we don’t but perceive, greenhouse gasoline elimination, and grid infrastructure that has been deferred for many years. “This could unleash us to the subsequent stage of creativity,” he stated.
Shevelenko made a extra pragmatic level: the entry-level job could also be disappearing, however the skill to launch one thing independently has by no means been extra accessible. “(For) anyone who has Perplexity Pc . . . the constraint is your individual curiosity and company.”
Younis drew the sharpest distinction between data work and bodily labor. He pointed to the truth that the common American farmer is 58 years previous and that labor shortages in mining, long-haul trucking, and agriculture are continual and rising — not as a result of wages are too low, however as a result of folks don’t need these jobs. In these domains, bodily AI isn’t displacing prepared employees. It’s filling a void that already exists and appears solely to deepen from right here.
While you buy by means of hyperlinks in our articles, we may earn a small commission. This doesn’t have an effect on our editorial independence.
