Three sample briefs, made only from what leaders said on air

the models have shifted from being the product to being a part of the product

We think that because it’s served up on one model, there’s a compounding element of advantage that comes from that.

I think the majority of actions on the internet will be taken by agents. Models are going to become more cheaper and faster at rates that are, you know, quite incredible.

the most important one that the world hasn’t really thought about yet is what it means for this to be a platform that everything integrates into

I think there’s something that belongs on a table. There is something that belongs in your pocket. And there is something that belongs on your body.

And then when Coding Agents started to work recently, we shut down things like Sora that we were also really excited about in the browser to really work on this.

They were just not as good as the most important thing to do, which was sort of push on the general capability of the intelligence.

We are not in the business of replicating the results of other labs, of kind of catching up to other labs in terms of benchmarks.

all researchers want to see products in the world. They have these incredible breakthroughs. They don’t want to just publish a paper.

if we had 3x to compute today, I could immediately utilize that very effectively.

we knew we’re going to launch GPT-5.5. People are going to love this model and we’re going to run out of compute.

I mean, I definitely badly undershot on the compute investments.

compute for us is not a cost center. It’s a profit center, right, when you deploy it in the product.

I am worried about the world’s compute buildout plans.

if that number were like 20 instead of 70, we should be much happier.

He is talking about the five-year cost of one gigawatt, which he puts at about $70 billion.

we are not a company that should be optimizing for free cash flow today because there’s just too much opportunity.

what CSPs do for us, in effect, is they shift CapEx into OpEx.

No one remembers who went first, Google or Yahoo, Lyft or Uber.

So imagine putting memory and context next to intent. You should have a very potent ad platform

we will pay our plugins that are popular and are seeing a lot of usage. They’re going to get part of the revenue share as well.

we’re not going to raise your electricity bills. We’re going to pay for our infrastructure and our power. It will not be the rate payer that has to pay.

sometimes you may have to say no to some of the demand, including some of the open AI demand, right?

People partner to get leverage, right? Because in some sense, it’s always the question of vertical integration

because of the Microsoft relationship, we were not allowed to invest in OpenAI. So that pushed us

We’re putting way too much pressure on two labs. I’m delighted to partner with them and support them.

embrace who you are and what you could be rather than like who others are

The OpenAI growth rate has been around 3 to 4x a year. The Anthropic growth rate has been around 10x a year.

I think it’s genius, actually, and I think they’re still going to sell a lot of OpenAI tokens into codex

when they went from being a non-profit to a for-profit, they should have done a make right.

The flagship models of Anthropic and OpenAI can cost $2 to $10 per 1 million tokens. Openweight models can cost around $0.20 per 1 million tokens.

They are going to need to raise at least $100 billion a year just to survive.

you have to agree not to criticize the company, basically.

And there’s been this evaporative cooling effect at OpenAI where year after year, there are rounds of selection.

even when it has been hugely unpopular in the AI safety community for us to push for iterative deployment

the models had reached a level of capability where they were able to find that exploit in our sandbox environment

we are not a company that should be optimizing for free cash flow today because there’s just too much opportunity.

even spending $220 billion in CapEx this year, we’re leaving a meaningful amount of demand on the table.

I anticipate that we’ll be short of capacity for the next couple of years.

We don’t see any drop. We see no drop in demand at all.

you look at your demand signals, you look at your committed revenue opportunities that you’re getting in the forms of contracts

But even with all of that supply coming on, it is hard for the companies in our portfolio to find compute even well into ‘27.

You can actually look at the demand in aggregate pretty well and predict based on monthly growth rates, which have been about 30%, 40% month-to-month

if you just kind of look at our compute, people see a couple of labs, but I see more than that. I see five, six labs.

We’re putting way too much pressure on two labs. I’m delighted to partner with them and support them.

And those really large training clusters are largely being bought by a small number of customers.

if AI works overall, plenty of other people want all our compute. So they’re willing to pay.

if you overbuild compute, then you don’t have any, like, margins other than reselling it.

And the bottleneck is less electrons. For us, it’s more of powered shell, right, or, like, delivered data center capacity.

And often even the big sites unfold in 50 megawatt blocks. 50 megawatt blocks is 18 months.

the bottleneck has moved from chips to electrons.

Unless the entire supply chain, starting from the equipment suppliers, their suppliers, their sub-suppliers, unless everybody produces more, you can’t increase the output collectively.

I can design the thing in three months, but I can’t actually get it into real silicon at scale for nine months.

until right now, I don’t think making memory has ever been a great business over a long period of time.

an experienced project manager is probably in highest demand I’ve ever seen throughout my career.

I think we’re going to be in this constrained environment for three to five years at least.

I anticipate that we’ll be short of capacity for the next couple of years.

The forward, it’s not just that the spot market is tight. If you want to cluster a year from now, you have to pay a lot of money.

My base case is that the projects that are fully permitted will get done in 2027

The reason that’s important is that that means they don’t have to go to the capital markets, right?

So, like, I don’t know if we’ve seen the peak of external financing, but I think we’re getting close.

And that’s why I was really like, it felt really powerful that literally the day after Microsoft reports their earnings, they also announce a $25 billion bond sale.

In the past two years, Oracle’s total debt has increased from $94 billion to $155 billion. The ratings agencies are not amused.

you can raise a billion dollars you can go out and get gpus for three billion dollars and only pay a billion dollars up front

The Frontier Labs, the Hyperscalers, the people who have investment-grade credit want to do the thing you’re doing, and will simply try to disintermediate you if they have the opportunity to.

is the biggest question the credit providers are going to say is well who’s the offtake? Because the reality is that’s actually who I’m underwriting

And you ain’t building no data center with 15% debt. Forget it. It just doesn’t pencil out anymore.

Their cost of capital, if you take their interest expense, multiply by four, divide by the debt, was more than 9

But if for whatever reason, the demand for AI just doesn’t materialize as much as we think, then compute prices fall dramatically, then all of these commitments are underwater.

the preponderance of the commitments are in future, they’re massive, and they’re all non-cancellable.

We see high teens, IRRs on compute capacity investment. We’re talking about a three to four year payback period.

CoreWeave says that its payback period is 2.5 years. Their typical deal is five years.

they’re getting sub 12-month payback periods, right, which is obviously a signal for demand.

And if you look at like Amazon, for example, the biggest builder of AI infrastructure, all of their AI infrastructure investments are profitable today.

Right now, if you have a B300, depending on what you’re running and what you’re doing, you can make somewhere between like $20 to $50 an hour on that B300.

The math didn’t work. You could have wrapped that fiber around the globe 10 times.

In 1999, bandwidth was $1,500 a megabit. And you couldn’t get it.

We literally could get into a debt bubble in the United States of America because there’s so much debt being taken out by all these, you know, hyperscalers.

they’re financing the mine and they own the bank that’s loaning the money to finance the mine to buy those shovels.

It was a good report where he estimated that U.S. data center capacity will be double the demand by the end of 2027.

The AI centers are a huge electoral issue, right, that elites are maybe not focused on as much as they should be.

if a local community, through their board of supervisors or whatever, votes down a data center, it’s dead on arrival.

It expresses a bet that there will be stringent regulation on building new data centers, that that capacity will be curtailed

Our data centers, we bring power and pay for all the power we bring to the grid.

The golf industry or the almond farming industry consume way more water than data centers do.

40,000 Americans applied. We’ve graduated 250 people. And we have a 90% retention rate.

what the data center industry, like their biggest problem is we have a marketing problem.

The real estate in space is, you know, infinite. It’s free. It’s free. You don’t have to pay for it.

being in sun-synchronous orbits, getting access to 24-7 energy kind of, you know, it’s a pretty good reason.

in space number one we don’t need permitted land so your biggest cost is gone we don’t need batteries and backup power because they’re 24-7 in the Sun

In space for each solar panel you get five times more power than you get on the ground because you can put them in a sun-synchronous orbit

I think that it is highly unlikely to be economically feasible or practical in the next decade or two.

So Falcon 9 is 10x short on scale and 5x short on cost.

The test is it does need to have broad impact across the economy in shifting the productivity curve

the country that is going to really win is going to be the one that can scale up broadly on AI

a rural Indian farmer was able to just go ask the bot to say, hey, find out about this subsidy program I heard about on television

The worst thing for any country to do in the name of sovereignty, if they’re off frontier, then that makes no sense because you’re falling behind.

So if you really think about composing that mindset, tool set, skill set, and data set

one of the dictums I have is don’t use frontier models for non-frontier problems

if you are high ARPU, low usage, then you have a little bit of a problem. But if you are, we are the exact opposite.

the most dangerous times are when you feel like you’re very successful because you start forgetting what made you successful.

I do believe the models are getting commoditized.

the future of every company will be some unique foundation model, so to speak, that they have that represents their knowledge

you need to be able to identify it clearly as something that you own, you control, you created, and you’ve put in place a system to have that compound.

Your data is your context. You feed the model. You collect, in fact, these traces or trajectories of how work gets done between humans and agents inside the enterprise.

For this to be stable, more of that surplus that gets created by use of intelligence to remain at the firm.

And then at work, I would love to have a chief of staff, a researcher, a consultant working with me on everything I do.

I delegate to a bunch of agents every morning. Let’s say I get up in the morning and I say to agents, do this work for me.

We have 200,000 employees and we have, let’s say, 2 million agents or 20 million agents all in a loop.

I first need to know what are these agents? What are they doing? What are their reasoning traces?

just like how we are test coverage, we now have this new concept called cognitive coverage

We know that the chain of thought itself can be unfaithful.

It’s not a supply issue of chips. It’s actually the fact that I don’t have warm shells to plug into.

Sam Altman was on the same episode.

the lesson is that compute or log of compute is intelligence.

I’m surprised that there isn’t a distributed model architecture at runtime, not for distributed training, right?

And if that is going to be true, then the question is, what is the world GDP growth rate?

He is asking whether the economy can carry ever larger training runs.

this industry to which I belong needs to earn the social permission to consume energy because we’re doing good in the world

that community believes that this data center is great for them.

It’s a platform company and it’s a partner company. A platform is not stable without partners.

He is explaining why Microsoft invests in startups.

People partner to get leverage, right? Because in some sense, it’s always the question of vertical integration

He is describing Microsoft’s partnership with OpenAI.

we want to make sure that we can exceed the teacher in order to set the frontier ourselves over the next few years

we now are able to manufacture and ship a chip that is 30% cheaper than a GB200 inside of our own clusters.

I’ve learned that optimism is a renewable resource.

She is describing what she learned from Nadella.

we find that like about 40% of our queries each week are health related.

you don’t want to be shipping your org chart when you are building your agents.

They’re all growing in the high teens and Microsoft is not getting the credit for it.

The reason that’s important is that that means they don’t have to go to the capital markets, right?

If Microsoft actually eliminates 27% of its workforce, like they’re suggesting, then their total headcount will actually be lower than what it was before the pandemic.

And that’s why I was really like, it felt really powerful that literally the day after Microsoft reports their earnings, they also announce a $25 billion bond sale.

because of the Microsoft relationship, we were not allowed to invest in OpenAI. So that pushed us

if you do everything, you will lose to Microsoft or Google who does everything but also has like trillions of dollars

It turned out being ourselves was like such a counter-positioning advantage against Microsoft.

I think the ROI right now is very mixed. I think there is heavy investment in this technology by enterprise, but a lot of this is experimental.