I was able to tour $IREN ‘s Childress site last Friday to talk about BESS.
After seeing in person, I can confirm:
⚡️The scale and size of their operations are impressive. There were over 400 people on site working that day
⚡️Data centers are very high-end. Squeaky clean, very well put together.
⚡️IREN looks to be on / ahead of schedule working towards 31 EH/s
⚡️Earthworks are already underway for blocks 4 (~150MW of something), block 5 (~150MW of something) and block 6 (~100MW of something)
Shoutout to the @IREN_Ltd team for hosting!
Allied demand for American nuclear fuel accelerates everything we're building here in the US.
Thank you to our partners at @ExImBankUS and NEDC for supporting us as we expand abroad.
📢 $NUAI just signed to acquire full ownership of its flagship #AI datacenter project in West Texas, buying out Sharon AI’s 50% stake in TCDC for $70M.
A major step forward in the Permian Basin.
Read the release: https://t.co/MZKlOhiD8A
$IREN is pleased to announce the signing of a $9.7bn AI Cloud contract with @Microsoft
Key details of the transaction:
- $9.7bn AI Cloud contract value
- 5-year average term
- 20% prepayment
- 200MW (IT load) data centers
- NVIDIA GB300 GPU deployments
Refer to the press release and accompanying presentation below for further information
Press Release: https://t.co/3pG9N2HfkF
Presentation: https://t.co/uyysnL9ffS
@mithcoons@TheKamaHsutra haha oh Matty. You either just lied or got mixed up claiming something that never happened, but again, hit me with the Matty C mic drop!
Oh bud. That post was you saying “congratulations to everyone who bought short dated 08/29 calls” and felt sort of sarcastic, like you were downplaying the momentum as a short term thing.
That was not you admitting you were wrong. I was trolling you by thanking you because I bought short dated calls, because I took the time to understand their business.
@mithcoons@TheKamaHsutra I don’t hold a grudge against you Matty boy. I like giving you shit about $IREN because you made some bad calls but won’t admit you were wrong - and always try to spin everything as a mic drop 🎤 win.
@ShashiGurufocus Tech seems interesting. I’ve never seen their tech in the field. I see a lot of warranty risk on new battery tech that isn’t battle tested. OEMs are tough businesses. I see better risk adjusted returns elsewhere which is why I don’t own EOSE.
$IREN's Next Step into Multi-Cloud
(Many of my post long but all are hand written. This one is my coup de grâce.)
Everyone is speculating on whether IREN’s next customer is Google, Meta or some other hyper-scale company. This is not the way to look at it. The first step to understanding what will happen is to understand Multi-Cloud. To explain Multi-Cloud, I’ll use an example none other than the Nebius-Shopify relationship. Shopify specifies exactly what their relationship is with NEbius: “we [Shopify] are partnering with GCP, as our main infra provider, and also engaging with neo-cloud providers such as Nebius to utilize large training clusters” (1). In other words, GCP is Shopify’s main cloud but Shopify is using Nebius for GPU intensive functions due to Nebius' lower GPU cost. Before any Nebius analyst jump up and say it’s not for lower cost, further down the Shopify blog post, Shopify says “Our tooling abstracts away from the underlying cloud provider, allowing us to move between vendors to get access to the latest GPUs as they're made available” (1). Shopify is a Tier 1 Tech Company as an E-commerce company but not a Tier 1 AI Research Lab, when Shopify says large training clusters, they mean training clusters are larger than inference but not the massive scale training cluster xAI, OpenAI, etc are building for their next big run. Now why doesn’t Shopify run everything on Nebius? Shopify needs GCP for user analytics, fraud detection, Spanner DB, customer data protection, Identity Services, Dataproc, etc. Let’s put it this way, hyperscalers are hyperscalers for a reason: the breath and depth of their software platform/services. Now why doesn’t Shopify use GCP for everything? GCP is “out” of GPU compute even though Google has already spent $75B in capex, or as Microsoft CEO Satya Nadella puts it, “I’m good for my $80B” (2). Note that GCP is not really "out" of GPU compute, but it’s got so much demand that GCP GPU time commands ridiculous prices. In other words, nothing is ever “out”, it just depends on the price. Hyperscalers are conscious in their capex in order to maintain their asset-light model and focus on high margins from IP/Software.
Why doesn’t Google just build more datacenter faster and take that free business away Nebius? First, Google is max allocated the capex that it’s investors are comfortable with. Second, Google already maxed out its DC expansion capacity and operations teams. While GPU datacenters are not nearly as complex as TSMC fabs, just like TSMC fabs are not copy and paste, GPU datacenter that have 99.8% uptime are not copy and paste. When you guarantee 99.8% uptime, you are guaranteeing the long tail of hardware bugs. You get hardware bugs that show up on one board and not anther. Some bugs only show up above certain hardware temperatures. Some networking bugs only show up above certain bandwidth and/or certain packet types. Uneven levels of hardware wear across heterogenous and homogenous chips. Inverse logic for poorly documented driver code from the antique ages. Tracing voltage fluctuations all the way down the schematic. God Forbid EMI. Google has already assigned all its teams to build and operate its own datacenters and part of the bottleneck is the training of new teams. Colocation gives Google one of the most important component of power but doesn’t solve all the bottlenecks. The building and more importantly operating of GPU datacenters for IaaS once you have power is not copy and paste! @brianfry01 knows! Thirdly and most importantly, Google doesn’t want to get stuck with all the current gen GPUs - multi-cloud distributes the risk! So what’s the value of IaaS? Why is Microsoft signing a deal with Nebius? By extending Azure compute on top of Nebius IaaS, Microsoft gets X PaaS revenue and pays Nebius Y IaaS revenue where X > Y. X-Y is a cashflow machine for Microsoft without the capex burden, GPU obsolecence risk, and Microsoft like Google cannot build GPU datacenters fast enough anyways.
The more important question is how can IREN participate in Multi-Cloud? $NBIS's Inference as a Service stack (3) is: DataOne (Power + DC Construction + DC Physical Operations) with Nebius (DC Design + IaaS + PaaS) = Full Inference Stack’s Value Chain. IREN’s official partner is TogetherAI (4) and TogetherAI’s Inference as a Service stack (5) will be: IREN (Power + DC Construction + DC Physical Operations + DC Design + IaaS) and TogetherAI (PaaS). Now why does inference need PaaS? Didn’t this Jim bro tell me bare metal GPU was the best? When you write inference code for an application, you need to optimize the GPU kernels, CUDA capture graph, and batching for that application. If you are T1 AI Lab, you tune this yourself, but if you are most other Software company working on a Software Application, you just want to focus on the Framework, API calls, DB Schema, fine-tuning your training model, but you don’t care how your inference is optimized as long as it’s optimize. The PaaS portion of “Inference as a Service” is a software service that fine-tunes the application’s inference and runs it on top of bare-metal GPUs. T1 AI Lab will tune it themselves and run on bare metal GPUs. In other words, PaaS is literally a service, not what inference runs on top of. TogetherAI+IREN = Nebius+DataOne and can serve as a AI Cloud in the Multi-Cloud ecosystem.
Now here’s the banger. It doesn’t matter if IREN works with TogetherAI or AWS or Microsoft, to IREN they all serve the same purpose: the PaaS layer. Now if AWS/Microsoft is uncertain about IREN’s capabilities, IREN will prove it in its partnership with TogetherAI. In fact TogetherAI needs IREN more than AWS/Microsoft because TogetherAI doesn’t have its power, DC operations, etc, so TogetherAI will give better margins to IREN than AWS/Microsoft. But once AWS/Microsoft sees that IREN can be trusted as Y in its X-Y cashflow machine, it will want IREN IaaS to fight for PaaS market share since PaaS is made by deploying code (aka copy and paste). But just as Nvidia keeps Neoclouds as leverage, it may make sense for IREN to partner with TogetherAI for better margins and better leverage in the relationship. IREN+TogetherAI will be in position to serve as Inference/Small Training Cluster Cloud in Multi-Cloud for companies ranging from startups up to large companies like Shopify. As IREN builds credibility, IREN’s IaaS may be able to partner other strong PaaS/SaaS companies like Databricks. I’ll note that $CRWV PaaS > TogetherAI PaaS which might be the factor to push IREN to consider working with hyperscalers or Databricks, etc. Furthermore, IREN can work with model development focused startups like HumeAI; although HumeAI doesn’t have a PaaS, HumeAI can optimize the GPU kernels, CUDA capture graph, and batching for its own application. This is why Tim Delcourt is in SF: to court the TogetherAI and HumeAI and work with the next generation AI companies.