To The Moon

A few weeks ago I wrote an article “The Second Life of a GPU.” In that post I discussed a (hypothetical) scenario where GPUs retained value past their (presumed) 4-5 year useful life. What impacts would this have on clouds / neoclouds or the broader compute complex? Could compute contracts be “re-contracted” after their useful life? All some version of the same question - is the useful life of these chips actually much longer than 4-5 years. Well, this week on Coreweave’s earnins call we got a great nugget related to this. “What we are seeing today is that the upside of recontracting is real as we remain largely sold out of prior generations of NVIDIA GPUs in addition to the current SKUs. So as our earlier generation fleets roll off their original contract, they offer the potential to deli

I strongly believe there will be 3, very large, markets that emerge in the “model / token” market: Frontier tokens (ie from the large labs like OpenAI and Anthropic) Vanilla open weight tokens (Deepseet, Moonshot, etc) Customized (RL / SFT) open weight tokens So much talk recently makes these seem zero sum… And of course, they do take share from each other. If open weight models never existed closed models would take it all! BUT - zero sum assumes the size of the pie is constant…which it isn’t. The market is growing much faster than any one group can steal market share from the other. So everyone grows! This won’t happen for ever. Eventually the market will mature and we’ll settle on a more “stable” market share. But for now (and for longer than people will expect), hypergrowth persists! I

The vast majority of earnings calls from public companies are pretty boring. But every now and then a CEO pulls back the curtain a bit and goes into a level of detail that I find fascinating. One of those calls was Jay Kreps from Confluent (on their Q1 ‘23 call) discussing the TCO advantages of using Confluent vs open source Kafka. On that call Jay went into a lot of detail about the differences between paying a cloud vendor vs hosting your own software. Jay’s one of the best infrastructure entrepreneur, so getting a peek into his brain like that was pretty cool. Yesterday was another moment where I really geeked out on an earnings call. This time it was Andy Jassy from $Amazon.com(AMZN)$ discussing the ROI on AI capex (which is clearly a very hot

I debated the title for this article a lot. The other option was “The World Isn’t Binary.” I’m still not exactly sure which title better describes this article, but I’ve found the principal of living with a positive sum mentality more impactful in my life so I went with that title. Anyway..on to the post! There are no shortage of big debates right now in AI. Open vs closed models. Nvidia vs custom silicon. Will chips have value after 5 years. Will AI kill software. God model vs fleet of specialized models. Are we in an AI capex bubble or not. Model layer vs application layer. Will models commoditize. The list goes on! These were just a few that came to mind… I won’t address all of these today, but the answer for most is “you’re not thinking big enough.” I’m clearly a hyper-optimistic perso

There’s a new open weights model out this week many are talking about - Kimi K3. It’s not technically open weights YET, but on their launch blog, they said" “The full model weights will be released by July 27, 2026.” At first pass, what stood out initially is the size of the model. K3 is a 2.8 trillion parameter MoE (16 of 896 experts active per token), which makes it the largest open model ever released (I think this is true, but haven’t fully fact checked everywhere). It’s ahead of DeepSeek V4-Pro at 1.6T (which I believe was the largest before K3). It has a 1M token context window, native multimodality, and seems to be built for agentic work. A lot of the launch materials / demos focused more on things like long-horizon coding, demos where it navigated massive repos, multi-hour autonomo

Alex Karp had a pretty spicy segment on CNBC last week talking about a lot of things (many of which took shots at the large AI labs). The part I want to focus on today was the concept he talked about of “owning your weights.” In summary: should enterprises “own” their models or rent them (ie buy from a large lab). Should you own your weights to give yourself more control, better cost management, maybe even better performance with with RL’s models, and ultimately remove a single point of failure (or reliance). This could come in the form of developing your own models, or RL’ing open weights models on your specific workflows / data. There’s a future version of this post that will focus more on “owning your harness” vs “renting a harness” from the large labs, but I’ll save that one for anothe

I think there are two big questions hitting AI markets right now: If $SpaceX(SPCX)$ (via xAI) and Meta are all of a sudden turning around and renting out their compute capacity, are we really in a compute crunch or is the system starting to fill up with excess? What does the rise of open source models and the end of tokenmaxing mean for Anthropic and OpenAI? On point 1 - SpaceX famously created ~$2.32B of monthly revenue by selling (ie renting out) about 450k of GPUs to Anthropic, Google and Reflection. Details below. SpaceX lands another computing deal, this time with Reflection, an open source model development company. $150m / month for GB300s. SpaceX the Neocloud! Deal 1 with Anthropic Colossus 1 and Colossus 2. Anthropic took all of Colossus

There’s a saying that will become more mainstream very soon. And that saying is “Time to Power.” Time to Power is starting to matter more in the infrastructure buildout. There are of course bottlenecks everywhere, but access to power is a big one. Why does it matter so much? Let’s look at a hypothetical data center build. These oftentimes will cost billions, or tens of billions, and the majority is financed with debt. As a debt provider the main question you’ll try and answer is “what is the likelihood the borrower will be able to pay back this loan, and under what timeline.” To answer that question, they’ll dig in on when the cash register will start ringing for the borrower (ie when the borrower will generate revenue). You could procure the chips, acquire the land, do everything necessar

$SpaceX(SPCX)$ lands another computing deal, this time with Reflection, an open source model development company. $150m / month for GB300s. SpaceX the Neocloud! Deal 1 with Anthropic Colossus 1 and Colossus 2. Anthropic took all of Colossus 1 $1.25b / month ~325k total chips, split roughly into 150k H100s, 50k H200s, 125k GB200s ~$5-$6 / hour blended Deal 2 with $Alphabet(GOOG)$ Colossus 2 $920m / month ~110k GB200s / GB300s (not clear which chip) ~$11-$12 / hour Deal 3 with Reflection Colossus 2 $150m / month ~18k GB300s Amount of chips not disclosed, but if we use the same hourly rate as the Google deal, the implied number of GB300s is ~18k In Summary $2.32B / month >$10 / hour for Blackwells (which

Workflows workflows workflows…I’ve probably written about this in the past (hard to keep track of everything I’ve written at this point, so I’m sure there will be some repetitive content!) but I wanted to circle back to it. For SaaS companies, the conventional wisdom was that you developed a moat if you built a system of record - a platform that stored data. If you owned / controlled the data, you had a moat! Data has gravity. True, but that wasn’t the real moat. The real moat was the hundreds of workflows that grabbed data from that system of record and then got work done. Sometimes those workflows originated from the system of record platform itself. Sometimes the workflows originated elsewhere, and one stop of the workflow grabbed data from the system of record. The real moat is owning