#xAI

0 today

May 21Thursday

Financial Times · TechnologyMay 21, 08:23

Anthropic on Track for First Profitable Quarter

Anthropic is on track to record its first profitable quarter ahead of OpenAI and xAI; the RSS snippet does not disclose the quarter, revenue, profit figure, or accounting basis.

Why it matters: HKR-H/K/R all pass: the FT claim reframes Anthropic’s business race against OpenAI and xAI. Missing quarter, revenue, and profit figures keeps it below P1.

May 5Tuesday

Synced · WeChatMay 5, 11:59

Massive Idle Cluster: Musk’s 550,000 Nvidia GPUs Are Only 11% Utilized

The Information says xAI’s roughly 550,000 Nvidia GPUs have only 11% MFU, equal to about 60,000 effective GPUs. The post cites HBM I/O, inter-server communication, training idle time, and software-stack inconsistency; Meta and Google are listed at 43% and 46%.

Why it matters: HKR-H/K/R all pass: the 550k-GPU versus 11% MFU contrast is strong, with concrete efficiency numbers and bottlenecks. This is high-signal infra reporting, not a model or product release, so it fits 78–84.

Take

xAI reportedly runs roughly 550,000 Nvidia GPUs at 11% MFU, equal to about 60,000 effective GPUs. If The Information’s number is right, the issue is not GPU scarcity. It is that xAI has not tamed large-scale training as an engineering system. A 550,000-GPU headline scares investors in one direction. An 11% MFU number scares systems engineers in the other direction.

The source body did not load. The WeChat page only returned an environment-verification screen. So the usable facts here come from the supplied summary: The Information cites 11% MFU for xAI, versus 43% for Meta and 46% for Google. It attributes the drag to HBM reads and writes, inter-server communication, training idle time, and inconsistent software stacks. That distinction matters. MFU is not the same as a casual GPU-utilization graph in nvidia-smi. It measures useful model FLOPs against theoretical hardware peak. In large-model training, memory bandwidth, collective communication, pipeline bubbles, checkpointing, data loading, and failure recovery all eat the numerator. An 11% MFU is not occasional idleness. It says the cluster rhythm is broken.

The Meta and Google comparison is brutal. xAI is roughly 4x lower by that metric. That is the gap between buying GPUs and making GPUs behave like one training computer. Google has spent years co-designing TPU systems with XLA, JAX/Pax, Pathways-era scheduling, and Gemini-scale operational lessons. Meta is not just a Llama publisher; it has poured years into PyTorch, FSDP, Megatron-derived stacks, datacenter networking, and large training operations. xAI’s Colossus narrative has emphasized build speed: Memphis, 100,000 GPUs, then much larger numbers. But training does not end when the InfiniBand cables are plugged in. At 550,000 GPUs, every small mismatch turns into downtime, tail latency, and idle accelerators.

I have one major reservation about the report: the summary does not disclose the measurement window. Pretraining, post-training, synthetic-data generation, inference, evaluation, and idle reserve capacity can produce very different MFU numbers. If 11% is a fleet-wide average across a mixed campus, it should not be treated as the efficiency of a single Grok pretraining run. If it is measured on the main training job, it is ugly. The article body, as provided, does not disclose model size, GPU mix, parallelism strategy, sampling period, failure rate, or whether all 550,000 GPUs belong to one training pool. Those details decide how hard the indictment is.

Even with that caveat, xAI does not get an easy pass. Meta at 43% and Google at 46% are not theoretical ceilings. Strong large-model training runs often target the 40% range for MFU, with higher numbers possible in narrower setups and lower numbers at huge scale. If xAI sits near 11%, its capex is not converting into iteration speed. Nvidia’s H100, H200, and B200 peak FLOPs look clean on a slide. HBM bandwidth, NVLink or InfiniBand topology, and communication overlap decide how much of that peak becomes training progress.

The listed bottlenecks sound very plausible. HBM pressure often shows up around attention, MoE routing, KV-heavy patterns, or bad batch and sequence parallelism choices. Inter-server communication points to all-reduce, all-to-all, pipeline scheduling, fabric congestion, and stragglers. MoE models make this nastier because token routing fragments the workload and stresses the network. If Grok leans heavily on MoE, that architectural choice can buy parameter scale while dumping complexity into the cluster.

Musk companies are very good at compressing physical build timelines. SpaceX and Tesla proved that repeatedly. AI training clusters are a different beast. Software debt directly taxes hardware spend, and it compounds fast. xAI can keep buying GPUs and keep presenting Colossus as the largest cluster story. But if MFU stays in the low teens, every new GPU mostly scales the inefficiency. If the next Grok release does not clearly close distance with Claude, Gemini, and OpenAI frontier models, this 11% number becomes the uncomfortable explanation: xAI did not underbuy hardware; it underbuilt the training system.

Feb 6Friday

Dwarkesh PatelFeb 6, 01:02

Elon Musk: “In 36 months, the cheapest place to put AI will be space”

Elon Musk predicts that in 30–36 months, space will become the cheapest place to deploy AI compute. He cites flat power growth, permitting bottlenecks, and roughly 5x better solar output in space without batteries; the interview does not disclose a cost model or validation data.

Why it matters: This clears the featured line as source-authority commentary: HKR-H comes from the stark 36-month space-cost claim, and HKR-R from the power bottleneck every AI infra team watches. HKR-K fails because the transcript gives heuristics, not a disclosed cost model or serviceability/

Take

Musk states a hard claim: within 30 to 36 months, space will become the cheapest place to deploy AI compute. He gives three inputs for that claim: terrestrial power growth is too slow, permitting is too slow, and solar in space is about 5x more productive with no batteries. That still skips the part that decides whether the claim survives contact with reality: total cost of ownership after you include GPUs, launch, thermal control, radiation hardening or mitigation, in-orbit assembly, redundancy, networking, and failure management. The interview itself surfaces the awkward counterpoint first: energy is only about 10% to 15% of data-center TCO. If that ratio is even roughly right, he is saving the smallest bucket while moving the most expensive bucket into the hardest place to service.

I’d split his argument into two layers. Layer one: terrestrial power is becoming the bottleneck for frontier AI. I mostly buy that. Over the last year, getting megawatts in the US has started to look like a queueing problem, not a capital problem. Utilities, transformers, interconnect studies, gas turbine backlogs, and local permitting are all slowing deployment. Microsoft, Google, and Amazon have all leaned harder into nuclear, gas, and behind-the-meter generation. I’m recalling the Microsoft-Constellation deal around Three Mile Island from memory, so I’d want to recheck details, but the strategic direction is clear. xAI’s own Colossus buildout already followed that pattern: secure generation first, then fill it with accelerators. Musk’s complaint about one-year interconnect studies is not crazy if you’ve ever talked to people building large campuses.

Layer two: therefore space becomes cheaper than Earth. I don’t buy that leap. A terrestrial power bottleneck does not automatically hand victory to orbital compute. You still need to clear at least four hurdles.

First, launch cost. This whole thesis needs Starship-class launch economics to work. The transcript gives no price per kilogram, no annual launch cadence, no orbital assembly plan, no replacement schedule for failed hardware. Without those numbers, “cheapest” is a slogan.

Second, thermal management. Ground cooling is hard, but space does not make heat free. You lose convective cooling and rely on radiative rejection. That means radiator area, mass, orientation, and more system complexity. AI clusters are heat machines. If you move them off-Earth, your cooling architecture becomes part of the payload budget.

Third, reliability. Musk says GPUs are quite reliable once you get past infant mortality. That is a reasonable statement for ground deployments. It is not enough for orbit. Single-event upsets, radiation damage, power instability, connector failures, and limited serviceability create a different failure model. You can design around some of that with redundancy and shielding, but then cost goes up again.

Fourth, networking. The interview does not explain how training data, checkpoints, and inference traffic move up and down, or what latency and bandwidth do to the economics. If he means batch training in orbit with minimal downlink, say that. If he means mainstream inference, the networking problem is front and center.

There’s also a sleight of hand here that I don’t love. Musk takes “a solar panel in space produces roughly 5x more useful power” and pulls it toward “AI is cheaper in space.” Those are not the same statement. Solar productivity is one subcomponent of the energy line item. AI TCO is dominated by far more than panel yield. The pattern we’ve actually seen over the last year is data centers chasing power on Earth, not power chasing GPUs into orbit. CoreWeave, Crusoe, Oracle, and xAI have all gone after places with available power, land, and cooling, even when the geography is awkward. The constraint is permits and supply chains, not physics alone.

Some outside context matters here. Space-based solar power is an old idea. NASA, JAXA, and ESA have all touched versions of it for decades. The sticking point was never “can you generate power in space.” It was “can the full system beat terrestrial alternatives on cost.” AI does add a new variable: a very high-value load that is hungry, time-sensitive, and willing to pay for scarce energy. That is real. But in my view it makes extreme terrestrial solutions more likely first: captive power, modular gas, pre-contracted SMRs, more liquid cooling, and tighter co-location with generation assets. It does not get you to “space is cheapest” on a 30- to 36-month clock.

Honestly, this reads like three company narratives stacked into one sentence. For xAI, it reframes the problem as energy scarcity rather than GPU scarcity. For SpaceX, it suggests a future demand pool much larger than launch and communications satellites. For Tesla, it leaves room for the AI6 chip line he name-checks. As strategy theater, it is elegant. As an engineering forecast, it is missing the spreadsheet.

If I had to place a bet, I’d bet on space handling communications, sensing, and some specialized edge inference earlier than people expect. I would not bet on orbit becoming the cheapest venue for large-scale AI compute in 36 months. The title gives a timetable. The body does not disclose the cost model. Without that model, this is still classic Musk: directionally provocative, numerically under-argued.

Jan 30Friday

Bloomberg TechnologyJan 30, 06:15

Elon Musk’s SpaceX Said to Consider Merger With Tesla or xAI

SpaceX is said to be weighing a merger with Tesla, with an alternative combination with xAI also under consideration. The post only says people familiar described internal consideration; it does not disclose structure, valuation, timing, or Musk’s preference. The key issue is consolidation path, not a completed deal.

Why it matters: Bloomberg's sourcing makes this worth tracking: the Musk-empire recombination angle lands HKR-H and the xAI capital/control question lands HKR-R. HKR-K fails because the story only confirms internal consideration; structure, valuation, timing and governance are still undisclosed.

Take

Bloomberg says SpaceX is weighing a merger with Tesla or, alternatively, xAI, but the public facts stop at “internal consideration.” There is no disclosed structure, valuation, exchange ratio, board process, or timing. My read is simple: don’t read this as a synergy story first. Read it as a control-and-capital story. Which entity absorbs which one determines who is subsidizing whom.

Start with the Tesla route. On paper, it is the cleanest Musk narrative: cars, robots, energy, autonomy, satellites, maybe even communications, all under one public-market shell. But the accounting and governance get ugly fast. SpaceX is still a capital-intensive business. Starship, Starlink, launch infrastructure, and manufacturing cycles consume cash. Tesla is publicly traded, so any related-party transaction would trigger brutal scrutiny on fairness, disclosure, and whether public shareholders are backstopping another Musk asset. There is an obvious historical reference here: Tesla’s 2016 SolarCity acquisition. That deal did not destroy Tesla, but it permanently trained the market to ask “bailout or synergy?” before it asks about strategic elegance. This story lands in that shadow immediately.

The xAI route is a different play. That would bundle AI compute demand, distribution, and a bigger financing base into one narrative. I can see why Musk would explore it. Frontier-model economics now hinge on sustained capex: GPUs, data centers, power contracts, and talent retention. xAI needs a durable capital story more than it needs another headline. SpaceX brings Starlink, launch infrastructure, and an asset base that investors already treat as strategically important. The catch is that aerospace and AI are valued on very different logic. Software-style multiples and private-market scarcity premiums do not blend neatly with launch cadence, satellite ARPU, or hardware depreciation. The article gives no financial detail, so I’m not going beyond the logic. Still, mixed-asset mergers often make financing look larger while making pricing messier.

I’m skeptical of the “empire consolidation creates efficiency” pitch. Over the last year, Musk already combined xAI and X around the idea of a model-data-distribution loop. If SpaceX is now entering the conversation, that tells me the previous consolidation did not fully solve the capital problem, or that the next spending cycle is much larger. Honestly, this looks more like a financing machine searching for a larger collateral base than businesses naturally converging.

There is also a governance asymmetry people gloss over. SpaceX investors are used to private-market opacity and founder control. Tesla shareholders are not, at least not to the same degree. If Tesla is the vehicle, every disclosure gap becomes a public-market problem. If xAI is the vehicle, the issue flips: investors can tolerate opacity longer, but valuation discipline gets weaker. Either way, the hard part is not “does the narrative fit?” It is “who gets marked up, who gets diluted, and who approves the mark?”

I’d watch for two concrete signals. First, do we see mention of independent directors, a special committee, or a fairness opinion? Without those, a Tesla-side transaction is hard to take seriously. Second, how is SpaceX valued against the absorbing entity? Until that math is disclosed, this is not a technology story. It is a governance discount story.

Dec 13, 2024Friday

OpenAI NewsDec 13, 08:00

Elon Musk wanted an OpenAI for-profit

OpenAI said on December 13, 2024 that Elon Musk pushed in 2017 to convert OpenAI into a for-profit and sought majority equity, absolute control, and the CEO role. The post includes a timeline and email excerpts, saying Musk formed “Open Artificial Intelligence Technologies, Inc.” on September 15, 2017, and that OpenAI rejected those terms. The real signal is the capital logic: the post says the team concluded in 2017 that AGI would need billions in compute, with Ilya Sutskever referencing hardware spend below $10B.

Why it matters: HKR-H/K/R all pass: the headline has a real reversal, and the post adds specific 2017 control demands plus concrete compute-cost claims. It stays at 80 because this is a one-sided OpenAI legal narrative, not an independently verified product or research release.