Yesterday, OpenAI announced a $38B deal with AWS, adding Amazon to its growing roster of cloud partners (Microsoft, Google, Oracle, CoreWeave). That brings OpenAI's total compute commitments to a mind-melting $1.4 trillion over the next few years.

On paper, this looks like the AI infrastructure arms race going into hyperdrive. When you add up all the previous announcements and news from earnings last week, Big Tech is on track to spend $400B on AI in 2026 alone. Everyone's buying GPUs like they're going out of style. Nvidia's getting rich ($5 Trilli and countin’), and Ed Zitron’s list of AI industry grievances is getting longer (based on his math, he says the AI industry will need $2 trillion by 2030 to make all this CAPEX worth it). 

And yet, in a recent podcast interview this week with Brad Gerstner, Microsoft CEO Satya Nadella dropped a bombshell: "My problem today... it's not a supply issue of chips. It's actually the fact that I don't have warm shells to plug into."

Translation = The bottleneck isn't chips. It's power.

Microsoft has racks of H100s sitting in warehouses collecting dust. Not because they don't want to use them. Because they literally cannot plug them in. The power infrastructure doesn't exist.

Think about that for a second. Every time Microsoft buys $50B of Nvidia GPUs, Wall Street celebrates it as "AI investment" and bids up both stocks. But if half those chips sit unpowered for 18 months, the ROI timeline collapses. You're paying data center construction costs and chip depreciation with zero revenue to offset it.

Below, we recap the top insights from the video, and then analyze what all this means for OpenAI, Microsoft, and yes, the OpenAI-AWS Deal.

‍

Top Takeaways From The Interview:

• Sam Altman called the OpenAI-Microsoft partnership "one of the great tech partnerships ever," crediting Satya Nadella's early conviction as essential to their success.
• OpenAI clarified its exclusivity deal with Microsoft: stateless APIs remain Azure-exclusive through 2030, but other products (open-source, Sora, agents, wearables) can be distributed elsewhere.
• The OpenAI nonprofit is already capitalized with $130 billion of OpenAI stock, making it one of the world's largest nonprofits, with the first $25 billion directed to health, AI security, and resilience.
• Microsoft's M365 Copilot has become "bigger than any suite" they've ever released in terms of adoption pace, according to Nadella.
• Azure could have grown faster than 39% if Microsoft wasn't compute-constrained, with Nadella confirming they're "not demand constrained, we're supply constrained."
• OpenAI scaled compute 10x in the last year, with Altman suggesting that if they had 10x more compute, revenue wouldn't be far from 10x growth.

• Altman revealed OpenAI's founding insight: they had "no idea at all how the tech was going to go" but had conviction in just "pushing on deep learning."
• The scaling law challenge: Altman admits it's "something closer to log of intelligence = log of compute" rather than the more favorable relationship they'd hoped for.
• Nadella identified the real bottleneck isn't chip supply but "power constraints and the time to build data centers" near that power: "I don't have warm shells to plug into."
• Compute demand follows the Jevons paradox for intelligence: as it becomes more efficient, demand and usage will explode exponentially.
• Software optimizations on the inference stack are "much more exponential" than Moore's Law improvements on hardware, according to Nadella.
• The architecture of SaaS is fundamentally changing as the "agent tier is replacing the old business logic tier," forcing a decoupling of data, logic, and UI.
• High ARPU, low usage SaaS businesses "have a little bit of a problem" in the AI era, while Microsoft's low ARPU, high usage model generates massive proprietary data graphs.
• Search economics were "magical" with fixed index costs, but each chat conversation now requires burning "a lot more GPU cycles," fundamentally changing internet economics.
• Enterprise AI monetization has a clear path forward: "Agents are the new seats" in the SaaS business model.
• The construction and supply chain of new data centers represents "the reindustrialization of the United States" according to Nadella.
• When Microsoft first invested $1B, Nadella admits he never thought "this is going to be the 100 bagger," but it became one.
• Both the revenue share and exclusivity with Microsoft end early if AGI is verified by an expert "jury," though Nadella believes "nobody's even close to getting to AGI."
• Altman dismisses concerns about OpenAI's $1.4T compute spend, stating "we're doing well more revenue than that" and the real risk is not having enough compute to generate revenue.
• Bill Gates was "rightfully so skeptical" of the initial investment until the GPT-4 demo, which he called the best demo since Xerox PARC.
• GitHub Copilot's smartest feature is its "auto mode," which intelligently chooses the right model for cost/performance based on the prompt.
• Microsoft has a $400 billion backlog (RPO) with a short 2-year average duration, giving them "high certainty" for their capex investments.
• Regarding the Colorado AI law, Altman stated "I literally don't know what we're supposed to do" to comply, highlighting the challenge of state-by-state regulation.
• Nadella warns that 50-state regulatory patchwork won't hurt big tech but will hurt "anyone starting a startup," potentially allowing the EU to set global standards.

Key Predictions for 2026

• AI coding agents will graduate from multi-hour tasks to "multi-day tasks" by 2026, according to Altman.
• Altman predicts AI will make a "novel scientific discovery in 2026," calling it "a wildly important thing" that represents "super intelligence in some sense."
• A new UI paradigm is emerging: "macro delegation and micro steering" will replace traditional interfaces, driving OpenAI's hardware development.
• Altman envisions a device that is "always with you" with "really good contextual awareness of your whole life."
• "Someday we will make an incredible consumer device" that can run GPT-5 or 6 models completely locally at low power draw, predicts Altman.
• Altman fears a scenario where breakthroughs allow "everybody can run like a personal AGI on their laptop" after building massive centralized infrastructure.
• "There will come a glut for sure," says Altman about compute, though timing is uncertain—could be 2-3 years or 5-6.
• If cheap new energy comes online at scale, "a lot of people are going to be extremely burned" with existing infrastructure contracts, warns Altman.

The Business of AI

• OpenAI plans "steeply" growing revenue from four key areas: ChatGPT, becoming an AI cloud, a consumer device business, and AI-automated science.
• Altman denies a specific IPO date, saying "I assume it will happen someday" but would prefer to fund the company with revenue growth.
• The "single most appealing thing" about an IPO would be giving retail investors an opportunity to own a stake, according to Altman.
• Microsoft realized it needed a compute fleet that was "fungible" across AI workloads, geographies, and generations to avoid getting stuck with old chips.
• Having royalty-free access to OpenAI IP is "kind of like having a frontier model for free" if you're a Microsoft shareholder, notes Nadella.
• Microsoft's high-margin business remains its 1st-party products: M365 Copilot, Security Copilot, and GitHub Copilot.
• Nadella clarified Microsoft's $13.5B investment was not booked as revenue; Azure revenue is "purely the consumption revenue of ChatGPT."
• "Chat conversations are new docs," which all feed into Microsoft's Graph to ground future agents, creating a powerful data flywheel.
• This is the "golden age of margin expansion" according to Nadella, as AI transforms productivity economics.
• Microsoft will grow headcount, but "that headcount we grow will grow with a lot more leverage" than pre-AI employees.
• Software agents will "take a whack at all of the IT backlog" and help create the "dream of evergreen software."

Breaking Down the Microsoft-OpenAI Deal:

What Microsoft Gets Exclusively (through 2032 or until AGI):

• OpenAI's leading models (GPT-5, GPT-6, etc.) via stateless APIs on Azure.
• Revenue share on all OpenAI revenues (estimated at ~15% based on the podcast discussion).
• 27% equity ownership in OpenAI on a fully diluted basis (worth ~$135B based on recent valuations).
• Royalty-free access to OpenAI's IP for seven years.
• Logo placement at the top of OpenAI newsletters and products.

What OpenAI Can Offer Through Other Providers:

• Open source models.
• Sora (video generation).
• Agents.
• Codecs.
• Wearables and consumer devices.
• Any other products and services.

Total Microsoft Investment:

• $13.5 billion in direct investment (training compute, not booked as Azure revenue)
• $250 billion additional compute commitment announced last week
• Plus the original ~$1 billion from 2019

How the Deal Ends Early:

• Both the exclusivity AND the revenue share terminate if AGI is verified by an independent expert panel
• Sam and Satya have to mutually agree on the expert jury
• The panel makes a "relatively quick decision" on whether AGI has been achieved

Key Quote from Sam: "If we had super intelligence tomorrow, we would still want Microsoft's help getting this product out into people's hands."

Key Quote from Satya: "Having royalty-free access all the way till seven more years gives us a lot of flexibility business model wise. It's kind of like having a frontier model for free if you're an MSFT shareholder."

Here's what Satya said about the app layer, ARPU, and commoditization:

On ARPU and the Winning Formula:

Satya's thesis: "One of the things I love about our Microsoft 365 offering is it's low ARPU, high usage... people are using it all the time creating lots and lots of data which is going into the graph."

The key insight: If you're high ARPU, low usage, you have a problem in the AI era. But if you're low ARPU, high usage (like M365, GitHub), AI becomes an accelerant because:

• Users generate tons of proprietary data.
• That data goes into your graph/repo.
• AI needs that data for grounding/context.
• You can charge MORE for the AI layer on top (M365 Copilot is "higher than any other thing that we sell").

He said: "Thanks to AI, we are seeing all-time highs in terms of data that's going into the graph or the repo... Chat conversations are new docs, they're all going into the graph."

On SaaS Architecture Changing:

Old SaaS: Data + Business Logic Tier + UI (all tightly coupled)

New SaaS: The agent tier is replacing the old business logic tier. AI "doesn't respect that coupling" - it forces you to decouple.

Brad's question: Won't more value accrue to the "token factory" (infrastructure/GPUs) and less to the software layer?

Satya's answer: There are TWO factories:

1. Token Factory = Infrastructure layer (running GPUs efficiently, max utilization, heterogeneous fleet).
   • This is the hyperscaler's job.
   • "If it was that simple, there would be more than three hyperscalers by now."
2. Agent Factory = Intelligent applications layer
   • "A SaaS application in the modern world is driving a business outcome"
   • "It knows how to most efficiently use the tokens to create some business value"
   • Example: GitHub Copilot's "auto mode" smartly chooses which model to use based on the prompt (code completion vs. task handoff)
   • This requires eval loops, data loops, feedback cycles

His conclusion: "The new SaaS applications are intelligent applications that are optimized for a set of evals and outcomes that know how to use the token factory's output most efficiently."

On "Nothing is a Commodity at Scale":

Brad brought up concerns about commoditization - everyone's fighting on price, margins should compress.

Satya's response: "At some level, everything is a commodity right? Compute, storage... everybody saying wow how can there be a margin except at scale nothing is a commodity."

Why Microsoft wins:

• Scale advantages on cost structure
• Supply chain efficiency
• Software efficiencies
• "When you have the biggest workload there is running on your cloud, that means not only are we going to learn faster on what it means to operate with scale, that means your cost structure is going to come down faster than anything else. And guess what? That'll make us price competitive."

On Search vs. Chat Economics:

Search was magical: Fixed cost (the index) that you amortize efficiently. Cost per search = fractions of a penny.

Chat is different: "Each chat... you have to burn a lot more GPU cycles both with the intent and the retrieval."

His take:

• Enterprise monetization is "much clearer",i.e. = "Agents are the new seats."
• Consumer monetization is "murky."
• "We're in the beginning of the cheese being a little moved in consumer economics."
• Need to discover what the ad unit will be ("agentic commerce or whatever").

Bottom line: The unit economics of chat will never match search's magic, but the total value to humanity could be "much much bigger."

So what does all this mean for the OpenAI-AWS deal?

OpenAI isn't just diversifying its cloud partners for fun. They're losing $11.5B per quarter with $1.4T in compute commitments and $13B in current revenue. They need every available data center with power—now.

Martin Peers at The Information points out the AWS deal is actually small ($38B vs. $250B for Microsoft, $300B for Oracle). His hot take? That might be a blessing for AWS. If OpenAI struggles to raise money for its commitments, Oracle—which is borrowing heavily to build data centers—is way more exposed.

The bigger picture: Investor Aakash Gupta argues the entire AI trade thesis just shifted. Power constraints mean "the old moat was model quality and algorithm improvements. The new moat is physical infrastructure and energy access."

In the podcast, Satya made it clear Microsoft is playing the long game. They're building a "fungible fleet" that works across AI workloads, geographies, and chip generations. Sometimes that means saying no to OpenAI's immediate demands.

Meanwhile, OpenAI is saying yes to everyone. When you've promised $1.4 trillion in spending, you don't have the luxury of being picky about where your compute comes from.

Here's where it gets interesting: Sam Altman addressed the "is there ever enough compute?" question in the podcast, and his answer reframes the entire demand discussion.

"You can talk about demand for energy at a certain price point, but you can't talk about demand for energy without talking about different demand at different price levels," Sam explained. "If the price of compute per unit of intelligence fell by a factor of 100 tomorrow, you would see usage go up by much more than 100."

He says there are AI use cases that make ZERO economic sense at today's prices but would explode at lower prices. This is classic Jevons paradox—as you make intelligence cheaper (tokens per dollar per watt, as Satya put it), demand doesn't just increase, it multiplies.

The challenge? The scaling law. Sam admitted: "Unfortunately, it's something closer to log of intelligence equals log of compute." Which means you need exponentially more compute to get linear intelligence improvements. Not great!

This explains why OpenAI is committing $1.4T in compute spending while losing $11.5B per quarter. They're betting that if they can drive down the cost per token fast enough, demand will explode before they run out of runway.

So the AI race isn't slowing down per say. It's just hitting a different speed limit—one measured in megawatts (or let's be real, gigawatts), not model parameters.

Now, in case you're curious (like we were) about kind of power deals were in the work, we spun up a Deep Research instance to look into this; here's what GPT-5 dug up for us: 

Timeline of U.S. Data Center Power Deals for AI (2019–2025)

2019–2021: Laying the Groundwork for AI rastructure

• 2021: Construction began on a 960 MW data center campus adjacent to the Susquehanna nuclear plant in Pennsylvania (originally developed by Talen Energy’s Cumulus unit). Amazon Web Services (AWS) later acquired this project in 2024, after the first 48 MW came online in early 2023. This “nuclear-adjacent” campus provides AWS with a large power-ready site for AI compute, though full utilization will roll out over several years.

2023: New Substations and Fusion Bets

• March 2023: Microsoft partnered with the Chelan County PUD in Washington to build a dedicated $86 million substation (Jumpoff Ridge) for its planned data center campus in Malaga, WA. The PUD began construction in mid-2023, with a ~40-month timeline to fully energize Microsoft’s three-building campus by 2026. Microsoft started commissioning the first 18 MW of this site by August 2025 using interim surplus hydropower.
• May 2023: Microsoft signed a landmark power purchase agreement with Helion Energy to buy 50 MW of fusion power from Helion’s first-of-a-kind fusion plant, expected to be operational by 2028. This fusion PPA – the first ever by a tech company – signals Microsoft’s willingness to invest early in novel energy sources to meet future data center demand.

2024: Nuclear Partnerships and Grid Constraints

• September 2024: Microsoft and Constellation Energy signed a 20-year agreement to restart the 835 MW Three Mile Island Unit 1 nuclear reactor in Pennsylvania (shut down in 2019) in order to power Microsoft’s cloud and AI data centers. The reactor – now renamed the Crane Clean Energy Center – was originally slated to reopen in 2028, but grid operators fast-tracked its interconnection, moving the target up to 2027.
• October 2024: AWS and Dominion Energy announced a memorandum of understanding to explore small modular reactors (SMRs) at Dominion’s North Anna nuclear station in Virginia. The companies plan to jointly evaluate developing new mini-reactors (potentially ~300–500 MW) on the site to support Virginia’s exploding data center load. Around the same time, AWS also inked deals for five advanced SMRs – partnering with Energy Northwest, X-energy, and Dominion – to deploy reactors in Washington and Virginia as part of its carbon-free energy strategy.
• November 2024: The Tennessee Valley Authority (TVA) approved a highly contested plan to supply 150 MW of power to Elon Musk’s xAI “Colossus” supercomputing center in Memphis. TVA’s board greenlit a new substation and grid feed (via local utility MLGW), enabling the repurposed factory site to draw the full 150 MW by late 2024.

2025: Trillion-Dollar Investments and Gigawatt-Scale Deals

• June 2025: Meta signed a 20-year PPA with Constellation to secure the entire output of the 1.1 GW Clinton nuclear power plant in Illinois. The Clinton Clean Energy Center was at risk of retiring in 2027, but under this deal Meta will offtake its carbon-free power through at least 2047.
• July 2025: Oracle pivoted to on-site generation by partnering with Bloom Energy to deploy solid-oxide fuel cells across Oracle Cloud data centers. Fuel cells can be deployed in ~90 days and run on natural gas or hydrogen. Oracle identified Digital Realty as a colocation partner for some sites.
• July 2025: OpenAI and Oracle launched Project Stargate, developing 4.5 GW of new capacity across at least seven U.S. states. Flagship site in Abilene, TX (built with Crusoe Energy) scales from 1.2 GW to 2 GW by 2026.
• July 2025: CoreWeave acquired ten data center sites from Core Scientific, securing 1.3 GW of capacity with a pipeline up to 2.3 GW. One new 300 MW site in Pennsylvania will deliver its first 100 MW by 2026.
• August 2025: Google signed a deal with NextEra Energy to restart the 622 MW Duane Arnold nuclear plant in Iowa. The plant, shut down in 2020, will resume operations by early 2029. Google also partners with TVA and Kairos Power on test reactors.
• September 2025: OpenAI, Oracle, and SoftBank revealed five more Stargate sites totaling ~7 GW across TX, NM, OH, and more. Nvidia joined with a $100 billion commitment to deploy 10 GW of capacity.
• October 2025: Oracle partnered with VoltaGrid to install 2.3 GW of gas-fired modular generators using Energy Transfer pipelines.
• October 2025: Google signed a 200 MW fusion PPA with Commonwealth Fusion Systems (CFS), with power expected from their Massachusetts plant in the early 2030s.
• Late 2025: Many power projects from 2019–2021 finally come online. Amazon’s renewables and Google’s 8 GW of clean energy contracts now contribute to capacity. Meta issued an RFP for up to 4 GW of new nuclear. As Sam Altman noted: power, not algorithms, may define the winners in AI.