Sweetwater, IREN’s crown jewel campus, has been long in the making and investors are still trying to figure out the vision. Sweetwater 1 is a 1.4 GW unicorn site that is already energized and just sitting there, uncontracted, in the tightest market in history, with a vision that’s yet to be made clear beyond it serving as a NVIDIA flagship DSX campus.
Could it be a trinity between $IREN + $NVDA + $PLTR? Here are the dots on the mosaic I see. All speculation and NFA / DYOR, but thinking through these potential threads highlights the unique value that each bring to the ecosystem.
- IREN management cares about optionality above all else. The 2.0 GW at Sweetwater 1 & 2 is the crown jewel of their portfolio and will be a significant capital investment. Picking the wrong partner would be catastrophic - how can they do that when it's not clear which model (ChatGPT, Claude, Grok, Gemini, open source, etc.) will win? The answer is simple: you don't sell (at least exclusively) to the model companies. You sell to enterprises / sovereigns. Is there evidence they're already moving in this direction? Yes. In 2025 monthly updates, Sweetwater was being marketed to whole of site tenants. And in a 2026 X Spaces, @danroberts0101 noted IREN was pushing back on ROFRs, which suggests single tenant is no longer the ideal state. Mgmt. has also made a number of references to targeting enterprises / sovereigns directly over time. TBD still if that's the intended Sweetwater customer mix or if hypers / labs are still in the running. But the long term strategy is clear: direct to enterprises / sovereigns is the goal, and it has to be if IREN wants to fulfill their long term vision of becoming a hyperscaler themselves.
- NVIDIA has already announced that Sweetwater will be a flagship DSX campus. Flagship is their word. I don't think we should take it lightly. It's unlikely that a flagship would be given to a single tenant - that's hardly a display of capabilities since it's a single audience. Which makes multi-tenant more likely and if NVIDIA really wants to demonstrate the efficiency and efficacy of DSX, what better way to do it than to go direct to enterprises and sovereigns. And this is especially true given the competition problem below - NVIDIA needs to continue to sell chips to all players in the ecosystem and the best way to do that is to demonstrate at massive scale that NVIDIA’s stack generates the highest revenue per token. Siloing that value add to one or a few customers doesn’t demonstrate that as massively as a diversified enterprise / sovereign campus.
- NVIDIA has a competition problem. Hyperscalers are making their own chips (e.g., $GOOG TPUs, $AMZN Trainiums, etc.) and now the labs are getting in the game too (e.g., OpenAI / $AVGO Jalapeno). Jensen’s already started the narrative shift: NVIDIA isn’t a chip company, they’re an AI infrastructure company. They’re already moving into adjacencies (robotics, self driving, etc.). And if hypers and labs are making their own chips, it only makes sense for NVIDIA to scale their own model. Enter Nemotron, NVIDIA’s open source model, that’s already competitive with Chinese models and several frontier models.
- U.S. open source has to win. We are in an existential moment in AI. Chinese open source is formidably capable and open source is impossible to stop. It’s data and it can be decentralized. But serious enterprises are going to struggle to adopt open source for fear of having to pre-scrub CCP backdoors before putting sensitive data into the model. The obvious solution to this: we need U.S. open source models. Nemotron is that.
- So if NVIDIA’s Nemotron is so capable, why haven’t they pushed it? In short, because they aren’t yet ready to be a hyperscaler. They don’t have the GW of power, land, and powered shells to stand up the compute needed to compete. And while NVIDIA has invested in $CRWV and $NBIS, neither of those two have nearly the amount of GW as IREN has, with ~6 GW of secured, grid connected, renewably sourced power in large scale (multi-hundred to 1GW+) campuses in favorable jurisdictions globally.
- Where does Palantir fit into this? Enterprises and sovereigns can’t optimize Nemotron out of the box. It must be harnessed to work within their specific organizations / use cases. NVIDIA and Palantir announced a partnership that creates an “intelligent engine” for running and improving Nemotron open models on Palantir’s platform. NVIDIA provides the open foundation models and compute. Palantir provides the secure wrapper, data governance, integration with proprietary data and workflows, customization tools, etc.
Result: organizations can train and customize Nemotron models on their own data in isolated environments, retain full ownership and control of the resulting models and IP, deploy them securely, and continuously improve them via feedback, without risks of proprietary insights leaking into closed frontier models. In short, Palantir turns NVIDIA’s open Nemotron models into a governed, secure, self-improving operational AI system for sensitive environments. NVIDIA supplies the models and compute, Palantir supplies the enterprise software stack, security architecture, and continuous improvement mechanisms.
No enterprise or sovereign wants their most sensitive data training Sam Altman’s or Dario Amodei’s next model that will be gate kept to a handful of chosen companies / government organizations prior to mass market release. This Palantir / NVIDIA partnership appears to be aimed at solving that exact problem.
- “But sir, Palantir procures compute from hyperscalers and like NVIDIA they don’t have their own datacenter campuses.” Two answers: (i) if you’re building a stack that specifically avoids having customer data on infrastructure owned by competing model companies (which includes the hypers), you need segregated compute, and (ii) there’s evidence that Palantir has been searching for a large scale site. Specifically, Palantir toured $FRMI’s Texas site back in Q4 2025. That site was initially targeting ~1 GW of power by the end of 2026, now pushed back to 2027. The only uncontracted campus in Texas with >1 GW of already grid connected power that I’m aware of is IREN’s Sweetwater (1.4 GW at Sweetwater 1, and 0.6 GW at Sweetwater 2). We also know that hyperscalers are turning away customers because they don’t have enough compute for themselves, with Microsoft and Google being recent examples. That dynamic could push Palantir and NVIDIA to need their own compute for this partnership.
Finally, let’s say Palantir + NVIDIA Nemotron is a killer app. Where are all these enterprises and sovereigns going to get the large scale compute to run it when compute is extremely scarce? Palantir doesn’t have that level of compute today and neither do enterprises - running AI cloud is infinitely harder than cloud storage. On-prem won't be the scaled solution.
- I listen to every @theallinpod podcast, and I listen closely, for one reason: it’s one of the most valuable AI infrastructure channel checks. There are enough tea leaves from recent pods that I want to connect them to this theory:
(i) IREN sponsors All-In - they’re connected with the all-in network, which includes NVIDIA and Palantir, and @GavinSBaker.
(ii) @chamath predicted the $SPCX / Anthropic compute deal before it happened. He had to have been hearing or seeing things in the background that made him think it was at least possible. He has a history of making predictions that come true in this sector and it’s likely informed speculation.
(iii) Chamath has his own >1 GW site and has made a number of recent comments on the value of large scale terrestrial datacenter campuses. He also noted that we need to find a way to get massive compute to open source to address the open source dynamics (Chinese open source competition + Sam / Dario training their models on customer data) described above. He could certainly be talking his own book, but it is possible that he’s heard rumors suggesting a large open source campus is coming together and wants to position his own site as a potential next-up.
(iv) In the latest pod, Gavin Baker joined and made it very clear that NVIDIA could compete in open source with Nemotron but to-date they haven’t pushed it. @Jason asked him why and Gavin clammed up with Jason noting that it was clear Gavin couldn’t say more. It’s possible that was referring to the NVIDIA / Palantir partnership announced this week. But it was still an odd moment that gave me pause to wonder if there’s something bigger brewing behind the scenes.
- Why could there be something bigger brewing behind the scenes? In short, because building a >1 GW campus takes an extraordinary amount of capital, and there’s no way to secure that amount of capital without customer contracts. When Chamath made his comments about the U.S. needing large scale open source compute, the #1 headwind he highlighted was the capital need. Here’s the problem: if you’re building the IREN / NVDA / PLTR trinity campus at Sweetwater I’ve described above (where IREN brings the power, land, shells, NVDA brings the compute / Nemotron, and Palantir brings the harness and enterprise / sovereign customers), there’s a huge gap: how do you raise the capital before the customer contracts are signed to line up the financing? My best guess is you would need a JV or a large pre-arranged financing that probably includes a mix of debt & equity from all three, as well as potentially outside investors. Could Gavin Baker be part of that? Could @AltimeterCap (another friend of the All-In pod and large NVIDIA investor that’s been bull-posting of late on neoclouds)? Both Gavin Baker and Altimeter have spoken highly of neoclouds (Gavin at Sohn speaking of neoclouds generally and Altimeter actually mentioning IREN directly) so it’s clear that neoclouds are at least on the brain. It’s possible that this financing may also involve the U.S. government as it’s clear from the Anthropic saga in 2026 (first the DoW blowup in Q1 and now Mythos / Fable in Q2) that the U.S. government doesn’t want to be entirely reliant on frontier labs.
IREN has been in advanced discussions with counterparties for 12-24+ months now, and it’s unclear if the desired customer mix has changed post the NVIDIA partnership announcement. The best explanation I can think of for the lack of clarity into the specific Sweetwater customer vision is this: either they’re waiting for Vera Rubins and still continuing convos with hypers / labs (certainly possible), or the opportunity has gotten much bigger and financing is being arranged in advance of a massive announcement.
- I’m probably not right about any of this, but let’s say I am. Why then would IREN be advertising? Simple answer: for the same reason that NVIDIA will keep selling chips to hyperscalers and labs. The market for AI training and inference is so massive that every player is expanding the customers they serve and crossing into others’ territories in how they serve them (with today’s news of $META potentially selling compute being another recent example that further validates the value of cloud compute) - NVIDIA can’t stop selling chips entirely as it’s the majority of their business, hypers / labs still need the best chips, and NVIDIA’s chips are still the best.
For the same reason, IREN, who has the best and most GW, likely won’t stop serving hypers and labs and they likely aren’t going to bet all their GW on a single outcome. IREN has ~6 GW of secured power with their total portfolio (secured + unsecured pipeline) potentially in excess of 10 GW+. They need brand recognition (hence the marketing push) to serve enterprises and sovereigns directly and that can be separate from this hypothetical IREN / NVDA / PLTR partnership. Strategically, it’s in their best interests for IREN to remain a neutral provider.
It’s entirely possible that I could be right about something bigger brewing with IREN / NVDA / PLTR and that Sweetwater could serve that partnership as well as dedicate compute to hypers and/or labs, while continuing to pursue direct AI Cloud at other campuses (or even Sweetwater).
What’s clear is this: power, land, and powered shells are still incredibly scarce. Every channel check confirms this, with the squeeze now extending into 2028 and beyond. There’s every reason to believe that IREN can continue to do business with all players in the ecosystem. I’ve simply seen enough dots on the mosaic that I wanted to put this specific hypothesis in writing to both sharpen my own thinking and read the feedback / pushback from others on X. If you’ve read to the end, thank you.
