The hardware behind the software

Every AI breakthrough you read about runs on a specific kind of silicon. AI chips are processors built to do the matrix math at the heart of machine learning far faster and more efficiently than a general-purpose CPU. The category spans everything from the massive graphics processors that train frontier models to the tiny accelerators now embedded in phones and cars. For investors, the story of the last few years is that the companies making this silicon captured an outsized share of the AI boom’s value.

George Gilder, who runs the George Gilder Report and has spent five decades forecasting technology shifts, is now making a specific bet within that broad theme. His “Ambient AI” pitch argues the next wave of AI chips will not sit in data centers at all. They will be the low-power silicon that lets devices run AI locally, without a round trip to the cloud.

From the data center to the device

The first generation of AI chips was about brute force: giant GPUs in climate-controlled warehouses. The second generation is about placement. Running inference on the device cuts latency, keeps private data local, and eliminates the per-query cloud cost. Qualcomm has been building neural processing units into its mobile chips for years, and Apple now ships a dedicated neural engine in every recent iPhone. The direction is clear, even if the market is still deciding who wins the silicon layer.

That is where the Ambient AI pitch gets specific. The promo’s headline promises a company “40,000 times smaller than NVIDIA” that is “set to power all of Ambient AI.” The tease resolves to QuickLogic (QUIK), a San Jose fabless semiconductor company with about 51 employees and a roughly $192 million market cap. QuickLogic does not make giant training GPUs; it licenses embedded FPGA intellectual property so chip designers can bake a small, reprogrammable logic fabric into their own system-on-chip.

What QuickLogic actually sells

The distinction matters. A traditional FPGA is a standalone chip whose logic can be reconfigured in software after it ships. QuickLogic’s “eFPGA” approach is different: instead of selling a finished programmable chip, it licenses the IP so a customer integrates the programmable fabric directly, using less power and board space than a standalone FPGA would. For battery-powered and defense-grade edge devices, that is a real advantage. We explain the technology in more detail in our edge AI chips explainer.

The catch is scale. Fifty-one employees and a $192 million market cap make QuickLogic tiny relative to the opportunity it is attached to. AMD, which now owns Xilinx, and Lattice Semiconductor both field far larger programmable-logic teams, and the biggest edge-silicon buyers tend to favor suppliers with long qualification histories and deep support staff.

The fine print on the pitch

Two numbers in the promo deserve a second look. The “40,000 times smaller than NVIDIA” line is a size contrast dressed up as a thesis; it says nothing about QuickLogic’s valuation or moat. And the “$1 trillion wealth explosion” refers to the entire Ambient AI market, not the slice a microcap can capture. The “$6 stock” framing is also stale by several years: the shares traded near $8 to $9 when the ad was re-teased in March 2026.

None of this means the theme is wrong. On-device AI is real, and every major chipmaker is investing in it. But our semiconductor stocks explainer shows how crowded the silicon layer already is, and the honest question is whether a 51-person company is the specific winner. Gilder’s own letter has produced a Cloudflare, up more than 1,000%, right next to an Inseego, down about 98%, which is the dispersion a concentrated small-cap bet carries.

Training chips versus inference chips

One split worth knowing is the difference between training and inference. Training a model is a brute-force job that favors the giant data-center GPUs. Running a trained model, called inference, is a lighter task that can happen on much smaller, lower-power silicon. The Ambient AI thesis is essentially a bet that the inference side, the side that runs on devices, is about to grow faster than the training side. That distinction, more than any single company, is the real idea inside the pitch.

The bottom line

AI chips are the foundation of the entire AI buildout, and the next chapter really is moving off the cloud and onto devices. QuickLogic is a genuine, if tiny, participant in that chapter through its eFPGA licensing. The reasonable approach is to understand the company on its own fundamentals, its revenue growth, and its competition from AMD and Lattice Semiconductor, rather than on the strength of a trillion-dollar headline.

NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.