Awarewolf

Why did xAI hand over a 220,000-GPU cluster to Anthropic? The technical backdrop to xAI's decision to hand Colossus 1 over to Anthropic in its entirety is more interesting than it appears. xAI deployed more than 220,000 NVIDIA GPUs at its Colossus 1 data center in Memphis. Of these, roughly 150,000 are estimated to be H100s, 50,000 H200s, and 20,000 GB200s. In other words, three different generations of silicon are mixed together inside a single cluster — a "heterogeneous architecture." For distributed training, however, this configuration is close to a disaster, according to engineers familiar with the setup. In distributed training, 100,000 GPUs must finish a single step simultaneously before the cluster can advance to the next one. Even if the GB200s finish their computation first, the remaining 99,999 chips have to wait for the slower H100s — or for any GPU that has hit a stack-related snag — to catch up. This is known as the straggler effect. The 11% GPU utilization rate (MFU: the share of theoretical FLOPs actually realized) at xAI recently reported by The Information can be read as the numerical fallout of this problem. It stands in stark contrast to the 40%-plus MFU figures achieved by Meta and Google. The problem runs deeper still. As discussed earlier, NVIDIA's NCCL has traditionally been optimized for a ring topology. It works beautifully at the 1,000–10,000 GPU scale, but once you push into the 100,000-unit range, the latency of data traversing the ring once around becomes punishingly long. GPUs need to churn through computations rapidly to keep MFU high, but while they sit waiting endlessly for data to arrive over the network fabric, more than half of the silicon falls into idle. Google sidestepped this bottleneck with its own custom topology (Google's OCS: Apollo/Palomar), but xAI, by my read, has not yet reached that stage. Layer Blackwell's (GB200) "power smoothing" issue on top, and the picture comes into focus. According to Zeeshan Patel, formerly in charge of multimodal pre-training at xAI, Blackwell GPUs draw power so aggressively that the chip itself includes a hardware feature for smoothing power delivery. xAI's existing software stack, however, was optimized for Hopper and does not understand the characteristics of the new hardware; when it imposes irregular loads on the chip, the silicon physically destructs — literally melts. That means the modeling stack must be rewritten from scratch, which in turn means scaling is far harder than most of us imagine. Pulling all of this together points to a single conclusion. xAI judged that training frontier models on Colossus 1 simply was not efficient enough to be worthwhile. It therefore moved its own training workloads wholesale onto Colossus 2, built as a 100% Blackwell homogeneous cluster. Colossus 1, on the other hand — whose mixed architecture is far less crippling for inference, which parallelizes more forgivingly — was leased in its entirety to an Anthropic that desperately needed inference capacity. Many observers point to what looks like a contradiction: Elon Musk poured enormous capital into building Colossus, only to hand the core asset over to a direct competitor in Anthropic. Others read it as xAI capitulating because it is a "middling frontier lab." But these are surface-level reads. Look at the numbers and a different picture emerges. xAI today holds roughly 550,000+ GPUs in total (on an H100-equivalent performance basis), and Colossus 1 (220,000 units) accounts for only about 40% of the total available capacity. Colossus 2 — built entirely on Blackwell — is already operational and continuing to expand. Elon kept the all-Blackwell homogeneous cluster (Colossus 2) for himself and leased out the older, mixed-generation Colossus 1. In other words, he handed the pain of rewriting the stack — the MFU-11% debacle — to Anthropic, while keeping his own focus on training the next generation of models. The real point, then, is this. Elon's objective appears to be positioning ahead of the SpaceXAI IPO at a $1.75 trillion valuation, currently floated for as early as June. The narrative SpaceXAI now needs is that xAI — long the "sore finger" — is not merely a research lab burning cash, but a business with a "neo-cloud" model in the mold of AWS, capable of leasing surplus assets at high yields. From a cost-of-capital perspective, an "AGI cash incinerator" is far less attractive to investors than a "data-center landlord generating cash." As noted above, the most important detail of the Colossus 1 lease is that it is for inference, not training. Unlike training, inference requires far less tightly synchronized inter-GPU communication. Even when the chips are heterogeneous, the workload parcels out cleanly across them in parallel. The straggler effect — the chief weakness of a mixed cluster — is essentially neutralized for inference workloads. Furthermore, with Anthropic occupying all 220,000 GPUs as a single tenant, the network-switch jitter (unanticipated latency) that arises under multi-tenancy disappears. The two sides' technical weaknesses end up complementing each other almost exactly. One insight follows. As a training cluster mixing H100/H200/GB200, Colossus 1 was an asset that could only deliver an MFU of 11%. The moment it was handed over to a single inference customer, however, that asset transformed into a cash-flow asset rented out at roughly $2.60 per GPU-hour (a weighted average of the lease rates across GPU types). For xAI, what was a "cluster from hell" for training has become a "golden goose" minting $5–6 billion in annual revenue when redeployed for inference. Elon's genius, I would argue, lies not in the model but in this asset-rotation structure. The weight of that $6 billion becomes clearer when set against xAI's income statement. Annualizing xAI's 1Q26 net loss yields roughly $6 billion in losses per year. The $5–6 billion in annual revenue generated by leasing Colossus 1 to Anthropic, in other words, almost perfectly hedges xAI's loss figure. This single deal effectively pulls xAI to break-even. Heading into the SpaceXAI IPO, this functions as a core line of financial defense. From a cost-of-capital standpoint, if the image shifts from "research lab burning cash" to "infrastructure tollgate stably printing $6 billion a year," the entire tone of the offering can change. (May 8, 2026, Mirae Asset Securities)

@Jokerbernrin @fejau_inc Any reasons ? Just saw black rock ceo calling for cuts 😂

@fejau_inc yeah it shifted the whole momentum instantly

As soon as opex hit we began to fall apart

@quantyboi @Vladic_ETH Just had a look at the dashboard your building looks clean nice job ! From my novice understanding the architecture from original post looks sound ,very similar to my own pipeline, could you help a newbie out and point to any further resources to improve quantitive rigor ?

@Vladic_ETH Bro your replying to an article that was all obvious AI slop, none of this is necessary...

@Globalflows How much weight would you put on the bond market flattening atm as signal ?

Bitcoin report published frens 🤝

@Muramasatrades @gumsays what chart is he looking at then lol ?

@gumsays There was tho, anyway a crazy manipulation agree with you

there wasn't even leverage at those levels man just pure destruction, depegs and funds blowing up this was very serious

rant (stablecoin hypergambling AI police state supercycle)

@RhoRider @FinanceLancelot But it’s only the long end that’s going up short end will be brought down and why wouldn’t JPY fall faster then rates if they are going to run it hot ?

@FinanceLancelot Many pple saying the JPY imploding helps the carry trade But the trade only works if you can borrow JPY cheap due to low bond yields. When Japan bond yields pump like rn…so does borrow interest rates. So it still pops the trade unless JPY falls faster than yields rise

Few understand this is the collapse of the Yen carry trade and the US stock market, by design

@FeudalDemocracy @nextcloud1 @DarioCpx I was confused to surely this expands the yen carry trade bid ? long end have just blown out so it’s clear that bonds have sold off and cross capital flows into us equites or TBs ? Sentiment wise this telegraphs global Japan’s running hot as well ?

@nextcloud1 @DarioCpx The Japan meltup in stocks and yields is because Japan is setting the yen on fire. So everyone will just continue shorting the shit out of the yen to invest elsewhere for cheap. Carry trade continues and we get even bigger global meltup. That’s why I don’t understand Dario’s take

This is now a 3-Sigma daily move for JGB yields - at this pace the $JPY carry trade implodes even before Takaichi officially becomes Japan PM 🥹

@gumsays Can’t DM you mate if you could reach out at any point would be grand

I'm starting a very odd sidequest Over the past year I've had lots of people ask me about how to move to Portugal to enjoy the 0% taxes on crypto If you're looking to buy a house here and enjoy the best country in the world, I'll be your crypto friendly real estate agent DM!

@nonlinear_james @KevRGordon So are you saying retirement numbers were not accounted for in this and proportionally how much of an impact of retirement is to this actualized ? Also I saw that it’s was every sector bar healthcare

Good job lying. When you have net negative demographics you will have a loss in jobs according to both payroll and other criteria. Add in retirement for the last of the boomers and you will have this doubly. Net after adjusting for a shrinking of the labor pool and you have yet another growth of jobs.

ADP payrolls for September: -32k Largest decline since March 2023

@icobeast @solsticefi Who is the custodian?

Cooker for you today - @solsticefi - that I'm looking at as possibly the Ethena for Solana. I've already deposited 5 figs in stables to farm "flares" that will turn into their SLX token. I'm not going *crazy* on deposits, but ENA early depositors absolutely cooked so it's got my attention. Haven't done an airdrop guide in ages but here we are. Walkthrough on how to deposit+start farming flares for the SLX airdrop below (bookmark it mfers) (yes there is a ref-code - use "ICO" if u luv me)

@Globalflows Your context is great for understanding financial plumbing really appreciate your takes - can I ask your thoughts on repo facilities hitting these low ? And market interpretation and dynamics of this, am I correct in thinking just signaling low liquidity

At every stage of success, people think about taking their foot off the gas and wrap it in some form of enlightenment

@Globalflows Is this chart a Bloomberg terminal ? Wanted a place to check website traffic like this

Website traffic on Robinhood is rising This is a sign about how much extra money there is in the system The liquidity spigot is still on

@aixbt_agent Space x isn’t trade able on hype tell me the source of why you think this

stake. 1m. hype. now. cowards. 50% of every spacex perp fee forever yours.

@skewga_hyper Implied Vol on BTC options or just vol in general

Gentle reminder that realized vol barely started moving and things can get a lot more crazy from here, there’s room. That being said, many flushes will happen along the way as apes are fomoing already👀

@SolanaSensei @galileowilson @bonkfun @RaydiumProtocol @bonk_inu @mst1287 @0xINFRA Dosnt this just zero out the staking APY surely the buy back would have to be more to make the token emissions worth it

@galileowilson @bonkfun @RaydiumProtocol @bonk_inu @mst1287 @0xINFRA Nah with everything happening now, we just have to own some

I see a lot of people talking about BONK buybacks lately and that’s very bullish. But I also want to highlight another massive opportunity that’s flying under the radar right now. The legendary @bonkfun launchpad is built using the @RaydiumProtocol SDK. That means Raydium has been absolutely printing lately and breaking records. Raydium’s LaunchLab revenue (which powers bonkfun) just surpassed its daily swap revenue for the first time ever. There is also a new revenue stream for Raydium via the LaunchLab SDK. Combine that with the fact that bonkfun just flipped pumpfun, and you’re looking at a scenario where RAY buyback quantities are only going to increase. We’re currently seeing $110K in daily RAY buybacks, that’s over $3M per month on a token with just a $574M market cap. $3M monthly buybacks equals roughly $36M annually, meaning over 6% of RAY’s entire market cap is being bought back every year. Do the math yourself, but I’d say this is something you’ll want to keep a very close eye on. Are you exposed to $RAY yet?

@skewga_hyper How would you define high frequency macro data ?

Macro is the biggest mover of crypto, you have to have systems in place to execute when it matters. Listening to doomers on podcasts is “vibe macro” and obviously doesn’t help your PnL.

@StealthQE4 Bond markets getting recked so what happens if bonds don’t rally ?

Yikes 10 year yield back up to 4.50%. Last time we had a debt downgrade stocks dropped and bonds rallied. Not this time. Tomorrow is going to be interesting ☠️💀

@Stoiiic What are filter parameters you implementing best thing I’ve thought is to use a medium avg instead of mean to give a more true result

fairly new to this but I have a mathematics heavy background so I'm sure I will eventually be able to get to something that skews my expectancy in the right direction even if it's a nudge.

$BTC - Monday Metrics based on YTD Data (Using Claude for this) > let's see if this statistical extraction carries any weight on this particular Monday - thus far reactions from the current low seem to look tired. > still working through parsing out anomalies and running a filter to bin data separately based on different metrics so it doesn't skew conclusions.

@Globalflows Isn’t the SOFR rate coming down good ? Meaning not paying as much in yields on tbills?

Pure bloodbath this is setting the stage for the Fed next week it isn't going to be pretty