Your slogan is "one person, one GPU," and the mission is to make compute as ubiquitous as electricity. But GPUs are scarce and power-hungry in a way electricity isn't. What has to change for that vision to become reality?
If you want to make the dream of one person, one GPU a reality, which is similar to the dream Apple had at the start with one person, one computer, you have to recognize the scale involved. A GPU draws roughly a kilowatt of power, so that implies a 300-gigawatt-plus build-out for every person to have access to a full-time GPU. For background, I started Lambda in 2012 as a face and image recognition company. I'm a machine learning engineer, so my job early on was reading state-of-the-art research papers, and among them was the AlexNet paper, one of the first to use NVIDIA GPUs to train image recognition networks, published by Geoff Hinton, Ilya Sutskever, and Alex Krizhevsky. I was lucky to get early exposure to what was then an emerging academic field.
Lambda has since pivoted to being an infrastructure provider of those computational resources. I look at this as a continuation of both the computer industry and the transformations the US grid has undergone, whether the electrical appliance build-out of the 1950s or the spread of central air conditioning as the American Southwest became more populated. Those were all multi-gigawatt build-outs of our grid infrastructure, and I think the AI move is the next of those major phases.
You have a deep relationship with NVIDIA, leasing your chips back to them, yet they also supply your competitors. How do you deal with depending on a single supplier whose other customers compete with you?
We have a bunch of what we call superintelligence customers, and NVIDIA is one of them, alongside many hyperscalers and frontier AI research labs. The interesting thing is that NVIDIA applies these large models back into designing better chips, which is part of why they're such a large consumer of their own compute. A lot of the subcomponents within a modern GPU are increasingly AI co-designed.
In terms of how we deal with it, we're a company founded by machine learning researchers, here to support the industry and community, and NVIDIA recognizes that shared heritage because both companies have been so embedded in the community. At the end of the day, we're talking about the entire American grid, basically all cloud compute, the entire AI industry, so of course that's going to involve multiple parties, not just Lambda.
The NeoCloud thesis is built on a GPU shortage expected to last. But once GPUs are ubiquitous and your motto becomes reality, what does Lambda look like in a world where compute is abundant?
There are a few parts to that process, including the massive expansion of the grid and the growing recognition that GPUs are an immensely backable and securitizable asset class, in the same way credit and equity investors love to invest in franchise utility providers whose job is to construct a power plant. Traditional grid infrastructure is a great, reliable cash-flow generator, and it's becoming clear that investors are recognizing AI factories as this new asset class.
It's obvious with any infrastructure build, whether railroads, telecom, or the grid, that there's a pattern. Carlota Perez wrote a great book, Technological Revolutions and Financial Capital, studying the interplay between financial markets and the maturing of technology, how the financial maturity curve of the ecosystem evolves alongside the technology. That's the story here, compute becoming a utility, and for that to happen there's a lot to do.
You're targeting three gigawatts by 2030 and have said the industry can't hire fast enough. Where do you see the main bottleneck: chips, power, space, or talent?
Definitionally there's only one bottleneck at a time, but over a decade of planned capital expenditure you go through a bunch of different phases. Back in 2023 it was the chip-on-wafer-on-substrate, the CoWoS capacity TSMC didn't quite have to make the Hopper H100 GPUs. Now it's what we'd call land, power, and shell, the active data centers ready to intake servers.
If you unpack that, there are localized bottlenecks around labor, master electricians and plumbers and the trades in general, long lead times on gensets used for backup and behind-the-meter power, and long construction lead times that often have to be underwritten by an investment-grade or high-credit-quality offtake agreement. Now memory is becoming another bottleneck, with the memory players exerting much more pricing power. The point is that over a decade you see a bunch of different phases, and it's difficult to predict which becomes the singular next one, which is the trillion-dollar question everyone's chasing.
CoreWeave went public at around 20 times revenue. As a still-private company chasing gigawatt-scale AI factories, how do you weigh mega funding rounds against going public?
In general, as a company matures and its capital requirements increase, if you're a net issuer of credit or equity you want to issue into the largest possible market so your demand pool is bigger. US private equity markets have something like a 13 trillion dollar market cap, while public markets have over 100 trillion of equity value, so it always makes sense for companies to go into the public markets as they mature and grow.
That speaks to my earlier point about the maturity of the industry driving publicly traded credit securities for these massive infrastructure projects. You start with private topco financings and equipment financings, which move to SPV-securitized private credit, which then moves to, and we haven't quite seen this yet, public issuance of credit securities directly associated with a particular build-out, the way you see in the energy industry.
You started in 2012 as a facial recognition company. What did the pivot to GPU cloud teach you about spotting the next platform shift?
I've been through a handful of pivots. We started as a face recognition API, added a deep learning consulting firm to stay alive before we were venture-backed, and had a product called DreamScope in 2015 that turned photos into paintings using neural networks, an early image generation tool.
If you're at the frontier of a technology, you can spot trends before they reach the public mind.
What it taught me was patience, and to really believe the feelings you have when first exploring a technology. I remember reading a 2013 paper by Alex Graves, "Generating Sequences with Recurrent Neural Networks," with generated Wikipedia-style text and generated handwriting samples that could be guided from messy to clean. I looked at it and thought you could imagine a future where these models solve all kinds of sequence prediction tasks, text, handwriting, image generation. I'd spotted the trend early, but it took about a decade to propagate through the system.
As you watch this unfold, is it a net positive or negative for humanity? What are your hopes and fears?
Every technology has both. Some of the oldest myths humans tell are about the terrifying power of fire, Prometheus stealing it from the gods, opening Pandora's box, immense power we don't quite understand. Any technology has amazingly powerful positive effects alongside powerful negative ones, whether nuclear power, telecommunications, smartphones, social media, or the petrochemical industry. The things unlocked are immensely positive, and there are negative things that need to be taken care of.
Overall, the trend through civilization has been that industrialization and technology have had immensely uplifting effects on the extension and quality of human life. The average person in the bottom tenth percentile in the United States has access to refrigeration, frozen goods, and lighting, things literal royalty in the 1500s did not have. So the evidence for technology being a net positive is pretty strong.
There seems to be a qualitative difference with this technology, a kind of quantum jump, especially as we discuss whether it could soon be conscious. Would you still put AI at the same level as past technological developments?
I absolutely would, having had these exact conversations with everyone in the industry since 2013. In basic macroeconomics, the two input factors into an economy are capital and labor, K and L, and one argument is that this is a complete consolidation where labor is totally subsumed into capital. But the "this time is different" claim has so much evidence against it, with regard to GDP growth and energy density utilization within an economy.
I think the evidence shows this is a continuation of the technological revolutions that started with the advent of the computer in 1950. The same arguments were made for the loom or the cotton gin, the textile worker whose job is automated, for hundreds of years, and it's always biased toward the side that technology is generally a much more positive thing.
Why do you think people have that bias against new technology? Is it simply fear of the new?
The cognitive bias driving it is the fear of uncertainty. There's a lot of research showing people prefer a known slightly negative outcome to an unknown one, taking a known amount of pain like a pinch over something unknown happening to them. That's the bias that causes this conversation to recur.
