What This Session Is About
Most conversations about AI happen at the application layer — models, agents, prompts. Amir Hochman comes from the opposite end. His firm, DCX, is an engineering company that designs and delivers major data centers across finance, secure government facilities, and now the wave of AI compute. His vantage point is not that of a consumer of these facilities but of the people who have to build them under real physical and regulatory constraints.
The core message: AI demand is bending the entire data center industry out of shape. Global operational capacity was roughly 6.2 GW in 2024; on current trajectory it reaches close to 250 GW by 2029. That growth is almost entirely AI-driven — cloud demand barely moved between 2025 and 2026, but AI workloads are pushing power and land availability to their limits. And power, not chips, is now the binding constraint on how fast the industry can move.
How AI Rewrote the Data Center
You cannot build without secured power, and power is running out in the traditional hubs. The smallest new-age AI facility starts at 100 MW and runs up to 2 GW. For context, one megawatt powers roughly 2,000–3,000 US homes — so a 100 MW site draws the electricity of a small city. Nearly every site today is pre-leased before construction even starts, which means the demand you hear announced is demand that has not been built yet.
In 2024, 10 MW of capacity meant ~50,000 sq ft and racks averaging ~6 kW. By 2026, that same 10 MW fits in ~5,000 sq ft, and a single Blackwell Ultra (B300) rack pulls 80–150 kW. Nvidia is now talking about 1.1 MW per rack within a few years — meaning ~800 of the racks from six years ago collapse into a single rack. The building barely changes; what goes inside it changes by orders of magnitude.
Until about three years ago, cooling was air only — cold air pushed up through the floor, chips running around 38°C. Direct liquid cooling now takes water straight to the chip, and because the chip target is ~38°C you can use ordinary cooling water rather than aggressively chilled air. It is far more efficient per megawatt. The catch: 15–30% of every site still needs traditional air cooling — and that residual 15% is itself the size of an entire data center from three years ago.
Roughly $1 billion buys 100 MW — about $10M per megawatt, split ~$6–7M for equipment and ~$3–4M for land and build. Amir's firm recently bid a 564 MW site: a ~$5 billion project. That is why building 10 MW of your own capacity in the office is a non-starter — the infrastructure alone, before you run or power it, dwarfs the cost of simply renting compute.
Key Insights
- 01The demand is real, not a bubble — because we use it every day. Amir framed AI capacity as fundamentally different from a speculative build-out: generating a single image consumes roughly the energy of charging your phone from 0 to 100%. Billions of those actions happen daily. The open question is not whether the energy gets consumed, but who owns the facilities and where they get built.
- 02Edge and small distributed data centers are not coming for AI. LLM training and inference need extreme connectivity between GPUs — Nvidia's architectures assume ~250 meters between the core and the farthest GPU. Interconnect is already ~10% of cost in a large site; in a small site it balloons past 40%. So compute concentrates into massive facilities near the "big seven" (Oracle, Microsoft, Google, AWS…), not into edge nodes near users.
- 03Build-your-own rarely beats renting today. Five years ago IBM, HP, Google and Microsoft all ran their own data centers — then most of the market dropped them and moved to lease. With GPUs carrying ~50% duty cycles, 3–5 year refresh cycles, and maintenance that now needs plumbers as much as engineers, on-prem AI infrastructure only pencils out at very large scale. For 10–20 MW of need, renting from a neocloud is cheaper.
- 04Geography is being redrawn by power price and regulation. Virginia was the heart of the internet and is now running out of power. Texas absorbed the first wave (OpenAI), and end-user electricity prices there rose ~30% in two years because private grid operators sell to the highest bidder. Demand is spilling into Ohio, West Virginia, Kansas, and Missouri. Kansas City already hosts 45+ data centers with another ~3 GW planned; Google, Microsoft, and Meta are all buying land there.
- 05Operators are building behind the meter — including their own reactors. As grids cap new connections (New Jersey sits at ~1.05 GW and is limited to ~1.35 GW), the largest players are generating their own power: Meta and Bezos-backed projects developing nuclear for ~10 GW of clean data center capacity, and others drilling natural gas wells — anything to avoid competing for scarce grid capacity.
- 06None of the technology is new — the implementation is. IBM ran cold-plate direct-to-chip cooling in 1988; the Cray-2 used immersion cooling in 1985. What is new is doing it at gigawatt scale, under a 100% SLA target (designed to ~99.99%), against a labor market that does not yet have enough people to maintain what is being built.
Today we need more plumbers than AI programmers. It wasn't a mistake when Jensen said the most important job in the future will be plumbers — the facilities are growing too fast, and we don't have the people to maintain them yet.
From the Q&A
With so many LLMs moving to local models, is that reshaping how hyperscalers plan demand?
Not yet, and probably not for AI at scale. Because of the connectivity requirements between GPUs, there are no meaningful edge deployments — models get built in big facilities close to each other. Running a small local data center is very expensive per unit of capacity because the infrastructure cost doesn't scale down. You might see it eventually, but for now the economics push everything toward large, centralized sites.
A large US bank told me their AI spend hit ~$1B across a few departments. At what point does it make sense to build rather than rent?
There's no single sweet-spot number. But building your own means a much heavier cost of ownership: every time Nvidia ships new equipment, the cooling approach and mechanical design change, and re-tooling is roughly double the cost of maintaining a standard facility. Power consumption is similar whether you build or lease, and GPUs are getting more expensive, not cheaper. Below a very large scale, renting from a neocloud stays cheaper.
What are the main failure points — what actually fails most?
Two things. First, electricity: generator backup burns ~540 liters of diesel per megawatt per hour, so storage is manageable up to ~50 MW but becomes impossible to hold for a long outage at 200–300 MW. Second, spiky loads: AI workloads ramp up and down so fast that to the electrical equipment it can look like a fault circuit. That frequency problem is still largely unsolved. Beyond those, the facilities are very reliable.
If you were deploying fresh capital today, which layer would you back — generation and grid, cooling, prefab, chips, or operators?
Operations. There simply aren't enough companies with the manpower to run these facilities. Data centers today run at only ~60% average energy load because operators can't fully implement redundant capacity — getting to 100% utilization is both the environmental and the efficiency win, and it's a real business opportunity. Robots aren't there yet, so skilled operations and maintenance is where the shortage is most acute.
How is the industry building faster given the timelines?
Prefabricated, modular construction. Data center components get manufactured in Europe, Canada, or the US, then delivered and assembled on site — cutting roughly 40% off the schedule. As a benchmark: a 10 MW site in Israel took 13 months without prefab (the fastest ever built there), while a 140 MW site is now being finished in 16 months using prefab. Without it, delivering at today's scale and speed wouldn't be possible.