Latest / Minds, Bodies, and Terawatts / OpenAI Raises $10 Billion in New Funding Round

Transcript

OpenAI Raises $10 Billion in New Funding Round · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where the money flows faster than the electrons. Bloomberg reported on March 24th, 2026 that OpenAI is closing a ten billion dollar funding round led by MGX, Coatue, and Thrive Capital, with the deal set to finalize next week.
• 0:15
  [Guest] Ten billion dollars. That's roughly the GDP of Malta, flowing into a single AI company in a single round. And the lead investor, MGX, is the Abu Dhabi-backed tech fund that only launched in 2024. Sovereign wealth is now treating AI compute like it once treated oil reserves.
• 0:32
  [Host] So who are these backers and why does their mix matter? Coatue is a classic tech hedge fund, Thrive is a Silicon Valley venture firm, and then you've got a Middle Eastern sovereign vehicle. That's a strange dinner party.
• 0:45
  [Guest] It's a very deliberate coalition. MGX brings patient capital and access to cheap energy in the Gulf. Coatue brings public-market savvy for when OpenAI eventually goes public or restructures. Thrive, run by Josh Kushner, has been in OpenAI since earlier rounds and is doubling down. Each investor solves a different problem for Sam Altman.
• 1:04
  [Host] Where does this money actually go? Because OpenAI isn't short on revenue. They reportedly crossed four billion in annualized revenue recently.
• 1:11
  [Guest] Almost all of it goes into compute infrastructure. Training the next generation of models, expanding inference capacity, securing long-term energy contracts for data centers. Revenue covers operations, but frontier model development is a capital bonfire. OpenAI's last funding round was six and a half billion just months ago. They burned through that fast enough to come back for ten billion more.
• 1:33
  [Host] That's sixteen and a half billion in roughly six months. At what point does this start looking less like a tech company and more like a sovereign infrastructure project?
• 1:43
  [Guest] That's the right comparison. What other entities spend sixteen billion dollars in half a year? Countries building highway systems. National defense procurement. The scale of capital flowing into AI compute is now indistinguishable from state-level infrastructure spending, except it's being directed by a handful of private actors.
• 1:59
  [Host] Let's talk about who this puts pressure on. Google just committed massive capital to Gemini infrastructure. Anthropic raised billions. Does this round change the competitive math?
• 2:08
  [Guest] It widens the moat, but maybe not in the way people think. The raw model capability gap between OpenAI, Google, and Anthropic is narrowing. Where ten billion dollars matters is in inference at scale. Running models for hundreds of millions of users, keeping latency low, keeping costs per query manageable. That's an infrastructure war, and infrastructure wars favor whoever can outspend.
• 2:29
  [Host] What about the startups that aren't OpenAI? If you're Mistral or Cohere, watching ten billion flow to your biggest competitor must be terrifying.
• 2:37
  [Guest] The mid-tier squeeze is real. Investors look at this round and ask themselves why they'd back a company that has to rent compute from the same cloud providers building competing models. OpenAI can negotiate power purchase agreements directly with utilities. They can co-design chips with partners. A startup with two hundred million in funding simply cannot compete on infrastructure. The market is bifurcating into a few giants who own the compute stack and everyone else building applications on top.
• 3:05
  [Host] There's something else buried in this story. MGX as lead investor means Abu Dhabi is making a strategic bet that AI compute is the next hydrocarbon. They're literally diversifying from selling energy to buying the thing that consumes it.
• 3:18
  [Guest] And they're positioned on both sides of the equation. The UAE has cheap natural gas, massive solar potential, and now direct equity in the company that will consume enormous amounts of electricity. If their grid powers the data centers and their fund owns the AI company, they capture value at every link in the chain.
• 3:37
  [Host] What about the geopolitical angle? There's been real tension in Washington about Middle Eastern and Chinese investment in American AI. Does this round face regulatory scrutiny?
• 3:47
  [Guest] It will get reviewed. CFIUS has been increasingly active on AI deals. But MGX has been careful to structure its investments to pass regulatory muster, and OpenAI has strong relationships with the current administration. The bigger geopolitical story is that American AI leadership now depends on foreign capital to fund its compute buildout. That's a dependency that didn't exist two years ago.
• 4:08
  [Host] So the money is secured. But can you actually spend ten billion dollars on compute without running into a wall? Not a financial wall. A physical one.
• 4:17
  [Guest] That's the question nobody on the investor call wants to dwell on. You can write a ten billion dollar check in an afternoon. You cannot build a gigawatt-scale data center in an afternoon. Permitting alone takes years. Transformer lead times are eighteen to thirty-six months. And the electricity has to come from somewhere. OpenAI can have all the capital in the world, but if the grid in northern Virginia or central Texas can't deliver another five hundred megawatts, the servers sit dark.
• 4:46
  [Host] So the real bottleneck isn't money. It's megawatts. And there's a piece on unscarcity dot ai called "GPT-4 Training: 50,000 MWh and the AI Energy Crisis" that puts real numbers on this. The article breaks down GPT-4's training run at about fifty thousand megawatt hours, enough to power forty-six hundred American homes for a year. But the part that caught me was the inference math. The article argues that every ChatGPT query uses roughly ten times the electricity of a Google search.
• 5:14
  [Guest] I actually want to push back on that ten-times figure. It depends heavily on which model is serving the query and what hardware generation you're running. A simple GPT-3.5 completion on optimized inference chips might only be three to four times a Google search. A long chain-of-thought GPT-5 query could be fifty times. The article's own data shows this range, but the headline number flattens it. What's more useful from that piece is a different finding: that inference electricity consumption scales linearly with user growth but training costs scale with capability jumps. So this ten billion dollar round is buying both curves simultaneously, and they compound differently.
• 5:52
  [Host] That's a fair distinction. But even taking the conservative end of that range, if OpenAI goes from two hundred million weekly users to a billion, you're multiplying a permanent draw on the grid by five. The article's core argument is that training is a one-time spike but inference is a compounding load, and that compounding load is what grid planners aren't modeling correctly.
• 6:14
  [Guest] Right, and that connects directly to the funding story. Ten billion dollars buys a lot of GPUs. But each GPU needs continuous power for the life of its deployment, not just during a single training run. The article traces how inference demand at OpenAI's scale could add the equivalent of a mid-size city's electricity consumption every year of user growth. That's the number that should worry grid operators more than any training headline.
• 6:39
  [Host] There's also a timeline analysis on unscarcity dot ai called "When Does Post-Scarcity Arrive? Realistic Timeline 2025-2060" that I want to bring in here because it directly challenges the optimism embedded in this funding round. The piece maps out when various energy solutions reach commercial scale, and it puts fusion at the late 2040s on realistic projections, with next-gen nuclear fission, small modular reactors, not reaching meaningful grid contribution until the mid-2030s at the earliest.
• 7:05
  [Guest] I think the fusion timeline is probably right, but the piece may be too pessimistic on fission. There are SMR projects breaking ground now that could deliver power by 2030 or 2031. The article acknowledges that possibility but assigns it low probability based on historical nuclear construction delays. And honestly, looking at the track record, that skepticism is earned. But the key argument in the piece, and this is where it really lands for me, is that AI investment capital arrives on a quarterly funding cycle while energy infrastructure operates on a decade-plus construction cycle. That mismatch doesn't get solved by writing bigger checks.
• 7:41
  [Host] Which means we're in a squeeze. This ten billion dollar round will deploy over eighteen to twenty-four months. The energy to power what it builds won't fully materialize for five to ten years. Someone is going to eat that gap, whether it's ratepayers facing higher electricity costs, communities competing with data centers for grid capacity, or climate targets that get pushed back because gas plants run longer than planned.
• 8:05
  [Guest] And that's the workforce dimension too. More compute means faster model improvement, faster deployment into enterprise workflows, which means the timeline for AI handling tasks in legal, financial, and medical fields compresses. If you were thinking you had five years to adapt, this round suggests you might have three.
• 8:22
  [Host] This is what Patrick Day-glon frames in the book Unscarcity as the timing problem: the capital for cognition arrives now, the physical infrastructure to sustain it arrives later, and the people caught in between bear the cost of the gap. You can explore over a hundred research articles, listen to the audiobook, and read the full book at unscarcity dot ai. This has been Minds, Bodies, and Terawatts.