Unconventional Inc. taped out a chip at Taiwan Semiconductor Manufacturing Co. on June 1, claiming the device draws orders of magnitude less power than conventional AI silicon, the company's chief financial officer Ali Esfahani said at the Bank of America Private Tech Trailblazers Conference in Palo Alto, California.

The startup, founded a year ago by former Intel AI chief Naveen Rao, has raised about $540 million and grown from 14 employees in January to roughly 60. Rather than separating compute and memory, Unconventional AI computes with physics-based dynamics on standard semiconductor processes; the changing state of the system itself holds the memory, in the way neural pathways do in the brain.

Esfahani argued that the energy grid can only grow linearly, leaving today's GPU architectures unable to keep up. "The energy grid can only grow linearly, so today's architectures can't keep up," he said.

Who else is building vertical AI on custom silicon?

The Unconventional tape-out was the most explicit hardware bet at the conference, but several other participants disclosed custom or co-designed silicon strategies:

• Bear Robotics is pairing its 16,000 deployed autonomous mobile robots with Nvidia's Jetson Thor onboard compute. Foundation models that learn from a few hundred examples and code written with large language models have cut work that once took six months down to days, co-founder Bren Pierce said. Tactile hands remain the missing piece, at about $30,000 each.
• CloudWalk, a regulated Brazilian payments firm serving more than 10 million users, runs its Pierre and JIM.com agents on a private cluster of hundreds of Nvidia Blackwell GPUs. Agents now handle 99% of customer support, up from 65% 18 months ago. The company clears $2.7 million in revenue per employee.
• Hippocratic AI layers 31 models into its voice-agent platform: one runs the conversation, 30 supervise it. Six health-system investors supplied 6 million real patient calls for fine-tuning.
• Airwallex says it has stacked up more than 90 licenses, banking partnerships and card-network connections since founding in 2015; the company has raised about $960 million across its Series F, G and H rounds at a valuation of up to $11 billion and runs about $1.4 billion in annual revenue run rate.

Why is domain-specific data suddenly the story?

Speaking on the sidelines of the conference, theCUBE Research executive analyst John Furrier framed the shift as a recognition that foundation models have become interchangeable, pushing competitive advantage toward proprietary data and industry-specific workflows.

"The thing about AI is that specialized intelligence is a big story now," Furrier said. "It was a sleeper story about a year ago, but domain-specific data, these verticals have the necessary requirements because the GPUs want data, the systems want data, the intelligence comes out of the data, and that is the key."

Bloomreach, which runs about 100 models inside its Loomi AI engine trained on 7 billion consumer profiles, said its vertical models perform five to 10 times better than off-the-shelf LLMs. Customers using four AI agents have grown 23-fold in a year.

Are capital markets really tilting back toward chips?

Yes, according to JD Moriarty, vice chairman and managing director, global head of TMT equity capital markets at Bank of America. "The biggest change in the market is how much capital AI and robotics need, which means companies now go public at far greater scale," Moriarty said. Public investors are rewarding durable, outsized growth, and for now scale is beating differentiation, the CoreWeave IPO having set the template. In 2026, the bank is seeing more activity in hardware and semiconductors than in software.

Other vertical hardware plays discussed at the conference:

• All3 (construction robotics) closed a $25 million to $30 million seed round and is preparing a six-story co-living building on an 11-sided plot, where labor accounts for 55% to 60% of total construction cost.
• Harbinger Motors builds electric and hybrid medium-duty trucks at diesel price parity, with California parcel fleets saving roughly $30,000 a year on fuel. A partnership with American Rheinmetall targets ground autonomy for defense, and Harbinger says it doubles output capacity every year.

The combination of Unconventional AI's TSMC tape-out, an air-cooled fleet of Nvidia Blackwell GPUs running a regulated payments product and Bank of America's read on hardware-led IPO flow points to a 2026 in which vertical AI companies stake their moats on process nodes, fab access and proprietary training data rather than on model novelty alone.