EIF Speaker Series: ‘AI’s Real Buildout Has Barely Begun’ with NG Zhang

The first wave of AI infrastructure has been defined by a rush for GPUs, data center capacity and power announcements. But as the industry moves from experimentation toward real-world deployment, the harder question may be less about how much compute can be announced and more about how efficiently it can be powered, cooled, financed and operated.
That is where NG Zhang, founder and chief executive officer of Canaan (NASDAQ: CAN), sees the market moving. Canaan is best known for making the world’s first-ever bitcoin ASIC mining hardware, but Zhang argues that some of the same lessons from mining’s evolution — energy discipline, specialized chips, modular systems and operational efficiency — may become increasingly relevant to the next phase of AI infrastructure.
This conversation is part of TheEnergyMag’s EIF Spotlight series, which features speakers attending the Energy Investor Forum and their views on the macro questions defining the next investment cycle across energy, AI infrastructure and bitcoin mining.
Zhang’s answers point to a broader thesis: AI may have started as a software story, but its next phase will increasingly become a systems engineering story. He sees the market shifting from announced megawatts to productive megawatts, from general-purpose GPU capacity to more specialized inference systems, and from standalone bitcoin mining sites to more integrated energy-to-compute platforms.
Below is the full Q&A with NG Zhang of Canaan.
Q: What’s a contrarian bet on the AI future?
I am only half-contrarian here. I agree with the mainstream view that AI is the future, and I also agree that energy will become one of the defining constraints. For Canaan, that is one reason our AI/HPC strategy has two pillars: energy and computing systems.
On the energy side, our view is that future compute infrastructure will be constrained by compliant, low-cost, financeable and operable power. The ABC Projects are important because they are not conceptual pipeline projects. They are energized, operating assets that generate computing power and cash flow. At the same time, we are working on larger and more controllable power resource development in resource-rich markets close to core demand. Energy projects take time — permitting, land, interconnection, construction and operations all require patience — so we prefer to make fewer promises and build more real assets.
The more contrarian part is on computing systems. I think AI compute may follow a more complex version of the bitcoin mining hardware curve. Bitcoin mining started with CPUs, then GPUs and FPGAs, and eventually moved to ASICs because the workload became clear and efficiency mattered more than flexibility. AI is more complex because models and algorithms are still changing quickly. But as inference becomes the dominant and recurring workload, the path toward domain-specific, inference-optimized computing systems becomes much clearer.
Training is how intelligence is created. Inference is how intelligence is used, distributed and monetized. If compute is mostly training, AI is still in the development phase. When inference dominates, AI has become real infrastructure.
For AI to become a true utility, compute has to become much more abundant and much cheaper. Software optimization will help, but software alone cannot solve the physics. The industry will have to confront performance per watt, cooling, power delivery, deployment cost and system-level specialization. This is where Canaan’s history matters. We have spent more than a decade working on ASICs, mining machines, large-scale delivery and real-world operations. We are familiar with turning high-density computing equipment into standardized, modular, mass-producible and remotely manageable systems.
So my contrarian bet is this: AI started as a software story, but the next phase will increasingly become a hardware efficiency, energy and systems engineering story. The scarce resource will gradually shift from GPUs alone to compliant power, dispatchable loads, DSA-based AI computing systems and long-term operational capability.
Q: Where would you invest $10,000 right now? Pick one each for high beta, low beta and negative beta.
I spend almost all of my time building a company, not trading stocks, so I would answer this more as a capital allocation framework than as stock picking.
For high beta, I would put capital behind energy-to-compute platforms: companies that can control or access power, bring it online, and monetize it across bitcoin mining, AI/HPC and other compute workloads. The highest-beta version of that bet is actually where I spend my own time and energy every day.
For low beta, I would look at the scaled, contracted layer of the AI buildout: real power infrastructure, electrical equipment, substations, transformers, cooling, and companies actually deploying AI infrastructure under long-term demand. Public-market volatility can still be high, but fundamentally these businesses are less dependent on guessing which model or application wins.
For negative beta, I would look at anything that credibly reduces the cost of inference by an order of magnitude: domain-specific architectures, specialized accelerators, edge inference, model compression, better memory architectures and energy-aware compute systems. Those technologies can be negative beta to overbuilt, general-purpose GPU data center economics. But they are not negative for AI itself. They make AI cheaper, which expands the real market.
Q: What comes after the AI buildout?
I would challenge the premise a little bit: I do not think the real AI buildout has even happened yet. We are still in the early gold-rush phase — selling picks and shovels, securing land, announcing power, proving strategic positioning. That phase is important, but it is not the final form of AI infrastructure.
The real buildout will be much larger than most people imagine, because AI only becomes transformative when inference is everywhere and cheap enough to be embedded into daily life, enterprises, devices, robots and industrial systems. That requires much more compute, not less.
After the first wave, I think the industry moves into a rapid iteration cycle. The market will stop rewarding announced megawatts and start rewarding productive megawatts: utilization, power cost, uptime, cooling efficiency, revenue per watt, customer quality and upgrade cycles. Specialized hardware and specialized sites can create enormous efficiency gains, but you do not build a skyscraper from the top floor. The messy first wave creates the demand, the supply chain and the operating knowledge. Then the real optimization begins.
Personally, this is what makes the current period exciting. It feels like the early years of bitcoin mining, but with a much larger social and economic impact.
Q: How does the US break its energy bottleneck?
I do not think the U.S. energy situation is as weak as some people describe. The U.S. is an old industrial economy with deep energy resources, mature capital markets, private infrastructure developers and a long history of building large-scale systems. AI infrastructure is actually an opportunity for the U.S. to move its industrial base into the next dimension.
The bottleneck is not simply electricity generation. It is time-to-power: how quickly a project can secure interconnection, substations, transformers, transmission capacity, cooling, permits, financing and reliable operations. A gigawatt on a slide is not the same as a gigawatt available to compute.
The solution has to be layered. The U.S. needs more generation and transmission, but it also needs faster large-load interconnection rules, more private investment into grid-adjacent infrastructure, and more flexible loads. AI/HPC loads generally need higher reliability. Bitcoin mining can operate as a more flexible, interruptible layer. In a well-designed portfolio, some load can be firm to customers, while other load can be flexible to the grid.
Canaan is not trying to become a traditional power plant developer. Our focus is closer to the layer that turns available energy into operational compute: power landing, modular deployment, thermal management, mining operations, and eventually broader energy-integrated compute infrastructure. That is where private companies can move fast and add real value.
Q: Does Bitcoin have a future?
Yes. I do not think bitcoin’s future depends on short-term price forecasts. The more important question is whether bitcoin continues to expand its role as a neutral, scarce, global settlement asset. My answer is yes.
I often think about bitcoin as having two major future chapters. The first is still human and public participation. Even after ETFs, institutional adoption and previous all-time highs, I do not think the broader public has fully participated in bitcoin yet. Market cycles can be painful, but temporary price weakness does not settle the long-term question.
The second chapter is more speculative, but potentially much larger: the machine-native economy. If autonomous AI agents become real economic actors — buying services, allocating resources, paying for compute, negotiating with other agents — they will need a native digital cash system. Bank accounts and credit cards were designed for humans and institutions. Crypto is much closer to machine-native money, and bitcoin remains the most credible digital reserve asset in that world.
That is why I think the long-term relationship between AI compute and bitcoin mining is deeper than people realize. Both are about energy, computation and digital value. Over time, AI infrastructure and bitcoin mining may become more unified in time and space: sharing power resources, sites, thermal systems, capital infrastructure and even economic networks. Bitcoin mining will not look the same as it did in the last cycle, but bitcoin absolutely has a future.
Zhang will join other energy, compute and infrastructure leaders at the Energy Investor Forum, where these questions — how AI gets powered, how compute systems evolve, and what role bitcoin mining will play in the next phase of digital infrastructure — will be part of the broader discussion.
Join EIF to hear more from Zhang and other industry leaders on how the next wave of AI and energy infrastructure will be built, powered and financed.


