What the Grok 5 Thread Gets Right, and What It Doesn't

The claim, and how much of it the record carries

An X post from 18 August laid out a specific account of what SpaceXAI is building at Colossus 2: Grok 5 in training since January, roughly 220,000 Nvidia GPUs online and climbing, power draw near 1 gigawatt, a 6-to-10-trillion-parameter target, an arrival before the end of 2026, and — the detail doing the most work in the thread — a training run on “the full 25-year SpaceX engineering data set.”

Some of that checks out cleanly against TEXXR’s corpus. Some of it undershoots numbers the record already carries, which is unusual for a hype claim. And one specific detail — the proprietary engineering dataset — appears nowhere in roughly 150,000 indexed articles. That gap is worth stating as plainly as the parts that hold up.

What’s confirmed: the training run, not the timeline

xAI’s own Series E announcement — a $20 billion round, oversubscribed against a $15 billion target, with Nvidia and Valor participating — said on 7 January 2026 that “Grok 5 is currently in training.” That is a primary-source confirmation of the thread’s “training since at least January” claim, and it is the most recent official word on Grok 5’s status the corpus holds. Nothing since then updates it: no parameter count, no training-data description, no revised release date from xAI or SpaceXAI directly.

What the record does carry, in the same stretch, is a track record of slipping the date. The Information reported in November 2025 that Musk pushed Grok 5’s release to “Q1 sometime,” later than an end-of-2025 target he had set earlier. Q1 2026 passed with no release. Since then, Grok 4.5 shipped July 8 as SpaceXAI’s first model built with Cursor, and Grok 4.6 followed August 13, pricing at $2 per million input tokens and matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index. The thread’s framing — that these are the incremental releases running in front of a much larger jump — is consistent with the pattern: two numbered point-releases inside six weeks, and nothing from the company itself about Grok 5’s status since January. Bloomberg reported in July that xAI had been “slowed down by internal chaos” through the first half of 2026, with sources describing the company as only recently “turning a corner” under product lead Michael Nicolls — one account, from inside the company, of why the gap between January’s training announcement and any public update has run this long.

What’s understated: the compute build-out is bigger than the thread says

Here the thread’s numbers sit below, not above, what the record has already disclosed. The Wall Street Journal reported that the original Colossus cluster, built in 122 days, ran on 100,000 Nvidia GPUs as of November 2024. By February 2025, Grok-3 launched trained on 200,000 GPUs — xAI’s own stated figure, “10x” the compute behind Grok-2. That December, the Financial Times reported xAI’s plan to expand Colossus tenfold, toward more than a million GPUs.

By October 2025, the Journal reported xAI was spending $18 billion or more to acquire roughly 300,000 more Nvidia chips for Colossus 2 — on top of a total Musk had put at 550,000 chips that July. Two months later, Bloomberg reported Musk had bought a third building, “MACROHARDRR,” adjacent to Colossus 2, that would take the site’s training compute to “almost 2GW.” That figure is eight months old and roughly double the thread’s “1GW and climbing.” Whatever the number of GPUs actually online this week — and the corpus has no article confirming 220,000 specifically — the trajectory the record shows is one that had already passed the thread’s power estimate before the thread was posted.

The one recent power figure the corpus does carry supports the general shape without confirming the specific one: TechCrunch reported August 1 that SpaceXAI is moving Colossus off the 69 unpermitted gas turbines that drew EPA scrutiny and a civil-rights complaint over pollution in nearby Black neighborhoods — covered in more detail on the power wall — and onto a permanent 1.2GW natural gas plant, with the turbines not fully retired until July 2027. A 1.2GW permanent plant, arriving after a December disclosure of “almost 2GW” of training compute already planned, is the more precise read on where power draw at the site is actually headed.

What has no source anywhere in the record

The thread’s most specific and most consequential claim — that Grok 5 trains on “the full 25-year SpaceX engineering data set” — does not appear in any article this screen can find. Searched across headline and description text for every combination of xAI, Grok, and SpaceX with engineering data, telemetry, flight data, or rocket data, the corpus returns nothing. Neither Musk, xAI, nor SpaceXAI is on record anywhere in roughly 150,000 indexed articles describing proprietary engineering or flight data as a Grok training input.

That silence does not make the claim false. It could be a real detail from a venue this corpus’s Techmeme-sourced feed hasn’t indexed — a livestream, an internal memo, a podcast. But it is the single load-bearing idea in the thread: the reason a 6-to-10-trillion-parameter model trained by SpaceXAI would matter more than another large language model is specifically that it might carry proprietary, real-world engineering data no other lab has. On the current record, that detail is one account’s claim and nothing more. The corpus can confirm training is underway, and can independently source a compute build-out larger than the thread describes. It cannot confirm the one thing that would make this a genuinely different kind of model.

What to watch

• Whether xAI or SpaceXAI issues any public update on Grok 5 — parameter count, training data, or release date — that the corpus can pick up and test against these numbers. The last one is seven months old.
• Whether the 220,000-GPU figure or a comparable current count surfaces in reporting, against a disclosed trajectory that had already reached “almost 2GW” of compute by December 2025.
• Whether the 25-year-engineering-dataset claim gets corroborated by an outlet, retracted, or simply never mentioned again — the difference between those three outcomes is the whole story.

Sources