Summary
KR Sridhar, founder and CEO of Bloom Energy, walks Harry Stebbings through why the AI revolution has finally made his 25-year-old solid oxide fuel cell company a $93 billion darling, up 1,500% in a year on a $20 billion backlog. His central argument: electricity is the sole physical input to manufacturing intelligence, the grid was built for the mechanical age, and power must move to the edge in modular, solid-state units the way computing did. Take him seriously and you stop treating power as a commodity utility problem and start treating it as the constraining resource on which the entire AI economy, and eventually global equity, will be rebuilt.
1. Bloom's 2001 pitch deck already drew today's AI data center, so the 25-year wait was "when," never "if"
Sridhar's original slide to John Doerr at Kleiner Perkins in 2001 showed a data center powered by a Bloom box with waste heat providing the cooling, connected to nothing else, which is precisely the configuration hyperscalers are buying now. He says there was "not a single night" in 25 years he doubted the company would work, only when. The company survived multiple existential threats, including a post-financial-crisis collapse of its entirely US automotive supply chain when the Big Three failed, forcing a Plan B within months against a six-year Plan A. Conviction is a supply chain of its own.
2. Andy Grove's "what's wrong with you" moment taught Sridhar that empathy, not expertise, is the CEO's job
Early Bloom boxes worked in the lab but failed spectacularly in the field because engineering intent had never been translated into instructions a shop-floor technician could execute. Sridhar convened his board and manufacturing gurus for an all-nighter with binders of data, and Grove immediately cleared the room, refused to look at the binders, and asked three times "what's wrong with you," then told him: "None of us know your product better than you. If you're smart enough to design it, you'll be smart enough to figure it out." The lesson Sridhar generalized: walk the floor, sit with customers to learn their pain rather than pitch, and toast the technicians at every executive dinner because pride on the line is what separates good from great. Empathy is a design discipline, not a soft skill.
3. AI is a hockey stick on a hockey stick because it is the first time humans are manufacturing intelligence
Sridhar frames AI not as another tech cycle but as a step change on top of digitization itself, because "when was the last time any civilization said we have too much intelligence, let's stop?" He dismisses bubble talk by separating stock prices from infrastructure: bumps and pauses are inevitable, but the trajectory of manufactured intelligence keeps compounding. As intelligence becomes ubiquitous and abundant, the scarce and valuable asset flips to wisdom, empathy, and human connection, things a bot cannot deliver. Price the intelligence, but invest in what it cannot replicate.
4. In an AI factory, electricity is the entire input, so power at the edge is the real unlock
An AI data center takes only two inputs, data and electricity, and data is everywhere, which makes power the sole physical constraint on manufacturing intelligence. Sridhar argues that 150 years after Edison, electricity is finally being redesigned for the digital age: modular, solid-state, and local, mirroring how computing moved to the edge with the smartphone. Latency-sensitive workloads like driverless logistics or robotic surgery cannot tolerate power outages or hundreds of miles of poles and wires exposed to weather, cyber attack, or physical sabotage. Power at the edge is not an efficiency play, it is a lifeline.
5. Bloom's architecture mirrors a data center's, which is why Larry Ellison chose it and eBay chose it first in 2013
Sridhar walks through the physics: a single 500 MW turbine runs at low-90s annual availability because it must be maintained like a jet engine, forcing hyperscalers to overbuild with a second turbine as backup now that grids cannot backstop gigawatt loads. Bloom's 50 kW modules hot-swap like server blades, ramp up and down in milliseconds to match GPU load spikes (eliminating the need for buffering batteries), and scale Lego-style as a data center grows rather than requiring a fractional turbine. eBay/PayPal became Bloom's first mission-critical data center customer in 2013 in Utah because they refused to relocate; when Oracle later needed 50+ MW in Utah after a power system delay, Bloom contracted for 90 days and delivered in 55. Match the architecture of the load, not the legacy of the grid.
6. Bloom sells "designer electricity," and the right metric is value per token, not cost per kilowatt-hour
Sridhar reframes the whole procurement conversation: grid power is one-size-fits-all, forcing customers to bolt on gear, cooling, and redundancy that they must own, operate, and maintain, with all the associated inefficiency and failure modes. Bloom's tailored stack collapses the full vertical, cutting the copper, tradespeople, gas, and supporting supply chain needed per token generated. The customer's job is to maximize the value of tokens produced, not to minimize a line item. Buy the outcome, not the electron.
7. The bottleneck is no longer Bloom, it is everything else around the data center
Bloom's backlog sits around $20 billion against $2 billion in 2024 revenue in a $5.5 trillion global electricity market, and manufacturing capacity is scaling from 1 GW to over 2 GW by year-end with a continuous "analog dial" ramp thereafter. Sridhar says building a 2 GW greenfield data center takes 12 to 18 months minimum, stressed by copper, cooling, and skilled trades supply chains, whereas Bloom can stand up power faster than the shell can be built. The remaining constraints are the customer's construction timeline, permitting, and natural gas supply, not Bloom's ability to ship. When your product ships faster than your customer can pour concrete, you have decoupled from the bottleneck.
8. Energy sovereignty is the geopolitical prize, and buying Russian gas while funding Ukraine is the tell
Sridhar ranks sovereignty priorities as food, then energy, noting that big wars have been fought over water, food, and now energy, and that the free world (US, Canada, Australia, UAE, Qatar) already has enough natural gas to displace coal globally and starve adversaries of leverage. He calls it "illogical" that the West still buys Russian energy while funding the opposing side of that war, and argues distributed Bloom-style boxes let free-world gas produce free-world power anywhere on the planet. Long-term he sees every community, from Iceland to sub-Saharan Africa, generating from local wind, solar, or geothermal, bottling surplus as on-site hydrogen, and recycling it through fuel cells for baseload. Sovereignty is a supply chain choice you make today, not a treaty you sign tomorrow.
9. Edge power ends the migration to megacities and could redistribute humanity itself
Sridhar's most contrarian claim: throughout history, people concentrated wherever access concentrated (rivers, ports, railroads, highways), which is why parents leave idyllic villages for poor conditions in cities so their kids get access. Bring electricity to the edge and you break the last physical reason for that migration, since information and services can already reach anywhere digitally. He argues 10 billion people should not have to live in metropolises to get a shot, and democratized power flips geopolitics the way the cell phone flipped the landline. Distribute the electron, redistribute the population.
10. Government equity stakes in AI companies are the wrong fix; a wealth transfer to the transition generation is the right one
Asked about proposals for governments to take ownership stakes in AI leaders, Sridhar rejects it: US market leadership churns every decade (Microsoft, Google were both once called monopolies), and state equity would grant unfair access that strangles the startups with better ideas. But he concedes the core concern is real: every technology revolution ultimately creates more jobs, yet the transition generation absorbs collateral damage, and Silicon Valley just told coders to learn to code before telling them AI will code for them. The construct he endorses is political, not regulatory: skim a portion of the abundance created and cushion those caught mid-transition. Protect the people, not the incumbents.
Ask this episode