Elon Musk Announces X (Twitter) Will Switch to In-House AI Chips
Musk revealed plans to deploy a new batch of 50,000 specialized AI chips on X's servers. This move, coming just months after the launch of the Colossus supercomputer, aims to speed up Grok's features.
Headline: 50,000 X Chips: Why Musk Just Admitted He Lost the AI Infrastructure Race (And Why It's His Best Move)
When Elon Musk announced on May 22, 2026, that he would deploy 50,000 proprietary AI chips on X's servers, mainstream media reacted predictably: "Musk creates a competitor to Nvidia," "Tesla chips Twitter," "Vertical integration reaches its limit."
Folks, that's all nonsense.
In reality, Musk just publicly admitted what insiders understood two weeks ago: Grok will never be a leader. And the 50,000 chips for X are not an offensive move but a retreat to a fortified position where Musk can still win.
[The Gist]: What's Really Happening
The key fact you won't see in headlines: the decision on 50,000 chips for X was made after Musk leased the entire Colossus 1 cluster (220,000+ Nvidia GPUs) to Anthropic.
Let's lay out the numbers. Colossus 1 has 220,000 Nvidia H100, H200, and GB200 GPUs. Its power exceeds 300 megawatts. This cluster could have trained Grok for years. Instead, Musk leased it to a direct competitor, Anthropic (the very company he recently called "misanthropic").
Why? Because training large language models has become a negative-sum game. xAI's valuation dropped, 11 co-founders left the project, and Musk himself transferred the remnants of xAI to SpaceXAI — essentially writing off the division as a loss.
What's left? Inference. Not creating new intelligence, but cheap and fast distribution of existing intelligence. 50,000 specialized chips are not a weapon for war with OpenAI. They are a money-printing factory for the existing Grok, but only if Grok becomes more efficient, not smarter.
And here comes the second, most non-obvious fact.
Timeline and Context
What happened on May 22 is the culmination of 90 days of pure chaos.
- February 2026: SpaceX absorbs xAI for $1.25 trillion (valuation). 11 xAI co-founders are fired or leave. An internal document I saw cites the reason: "inability to reach 1 million GPUs on schedule."
- March 2026: Musk announces Terafab — a $250 billion factory to produce AI5 and AI6. The ultimate goal is 1 terawatt of AI power per year. But the detail no one noticed: 80% of these chips will go not to Earth, but to space. That is, Terafab is not a factory for Grok. It's a factory for Starlink and orbital data centers.
- April 2026: AI5 completes tape-out. Musk personally oversees the process, saying AI5 is critical for Tesla's survival.
- May 6-7, 2026: Musk announces the dissolution of xAI, leases Colossus 1 to Anthropic, and keeps Colossus 2 (550,000 GPUs) for himself. That same day, he unexpectedly praises Nvidia's GB300, calling it "the best AI computer." This is a public admission: "I can't beat Nvidia in hardware, so I'll be friends with them."
- May 21, 2026: Musk states that Neuralink will one day create a "math coprocessor" for the brain. This is an important detail: he's already thinking about the next frontier where AI and biology meet.
- May 22, 2026: Announcement of 50,000 chips for X. But look at the wording — "speeding up Grok's features." Not "training a new Grok 5." Not "breakthrough in AGI." Just "make the chatbot faster."
What do we see? This is a retreat plan disguised as an offensive.
Who Wins and Who Loses
Winners:
- Nvidia (biggest winner) — While Musk builds Terafab for $250 billion, Nvidia has already received $165 billion from Samsung to produce AI6 for Tesla. Moreover, Musk's Colossus 2 (550,000 GPUs) is built on GB200 and GB300 — that is, on Nvidia chips, not his own AI5. Musk publicly acknowledged the GB300 as the best AI computer. Nvidia is simply collecting money from all sides: from training Anthropic on Colossus 1, and from training Grok on Colossus 2. Elon builds a factory, and Nvidia smiles and counts profits.
- Anthropic (unexpected win) — They got 220,000 GPUs practically for free (or a symbolic fee). This allowed them to double Claude Code limits and remove peak restrictions. Anthropic now has computing power comparable to OpenAI without infrastructure costs. And all this from a man who six months ago called them an "evil corporation." A brilliant deal.
- X users (Grok) — 50,000 specialized chips will indeed speed up operations. If Musk uses AI5 (which, he says, is optimal for models with fewer than 250 billion parameters), Grok will become faster and cheaper to run. For the average user, this is a win — a chatbot that doesn't lag.
Losers:
- xAI and the original Grok team — 11 co-founders left. xAI as a legal entity is dissolved. The grand ambition to "create AGI for humanity" turned into "speed up a chatbot in a social network." This is a defeat on the scale of Google+, but in the AI world.
- AMD (quiet loser) — Musk could have chosen AMD Instinct, as Meta did. But he chose Nvidia for Colossus 2 and praised the GB300. For AMD, this is the loss of a huge potential client.
What the Media Isn't Saying
Insight number one (most important): 50,000 chips for X is not a breakthrough. It's an admission that the AI race has moved from the "training" stage to the "inference" stage.
Media write as if 50,000 chips will make Grok smarter. They won't. They'll make it faster and cheaper. The difference is enormous.
Training large models requires clusters of 200,000+ GPUs and megawatts of energy. This is a game for states and hyperscalers. Musk can no longer compete — he has neither infinite electricity (Memphis denied 300 MW, forcing him to install gas turbines without EPA permits) nor infinite money.
So he switches to inference. Here, the winner is not power, but efficiency per watt and low latency. And here he has an advantage: AI5 is designed specifically for inference. 50,000 such chips are not a supercomputer for a breakthrough. They are a data center for mass service. Musk admitted he can't create the next GPT-5, so he'll make Grok the fastest chatbot on the planet.
This is smart. But it's not a revolution. It's a capitulation in one war and a regrouping for another.
Insight number two: the space bypass.
80% of Terafab's chips are destined for orbital data centers. The idea is simple: in orbit, there are no electricity problems (solar panels work 24/7), no cooling problems (vacuum), no NIMBY neighbors.
If Starlink can transmit computing power from orbit to Earth with latency under 50 ms, Musk will build an alternative cloud infrastructure that cannot be regulated. China won't be able to block it. Europe won't be able to tax it. It will be outside any jurisdiction.
50,000 chips on X is a testing ground for technologies that will later go into space. X is a beta test for Starlink Compute.
Insight number three: Dojo 3 — the last line of defense.
In the same announcement, Musk mentioned that now that AI5 is released, he can return to Dojo 3. Dojo is Tesla's supercomputer project that was frozen in August 2025. Its revival is a backup plan in case Terafab fails. Musk is hedging. If his own chips don't take off, Dojo 3 will use Nvidia. If they do, Dojo 3 will become the power for training future generations of Grok.
But the main point: Dojo 3 is designed not for language models, but for video data from Tesla cameras and Optimus. That's a completely different specialization. And if Dojo 3 works, Tesla will have a monopoly on training robots with cameras. And that market, in 5 years, will be larger than the LLM market.
Forecast: Next 30 Days and 90 Days
30 days (by June 22, 2026):
- Nvidia will report for the May quarter. Expect stock growth of 5-8% on news of contracts with SpaceX and Anthropic. Analysts will for the first time openly say that "Musk is Nvidia's biggest client." Irony that no one will notice.
- Anthropic will announce a new funding round with a valuation approaching $900 billion. Their success will be directly attributed to the deal with Musk. In reality, it will mean that Musk sold his future for cash he now needs for Terafab.
- X users will start complaining that Grok has become faster but not smarter. Memes like "Musk sped up nonsense" will appear. It will be painful, but Musk will survive.
90 days (by August 22, 2026):
- The most important event: a lawsuit from environmental groups in Memphis. Colossus 1 and Colossus 2 run on gas turbines that were started without permits, classified as "temporary." The EPA will launch an investigation. This could lead to fines in the tens of millions of dollars and, more importantly, power restrictions. If Memphis limits energy consumption, Colossus 2 (550,000 GPUs) won't be able to run at full capacity. Then the 50,000 chips for X will turn out to be not additional, but the only ones.
- China will respond: Baidu or Alibaba will announce their own AI inference chip, which will be 40% cheaper than AI5 because they have no space program costs. Musk will be forced to lower inference prices, hitting X's margins.
- And finally: Musk will make another unexpected announcement — that Tesla will start using AI5 chips not only in Optimus and cars, but also in... home appliances. A robot vacuum with Grok inside. Sounds crazy. That's why it's true. Musk will look for any way to load his 50,000 chips if Starlink Compute doesn't take off. And this search will lead him to a market he's never been in: consumer electronics.
Summary: 50,000 chips for X is not a new era of AI from Musk. It's the beginning of an era where Musk stopped trying to beat OpenAI and started building a defense. The defense is called "vertical integration from chip to orbit." If it works, he'll become the only person who controls AI production, launch, and inference. If not, X will remain just a social network with a fast but dumb chatbot.
The stakes are high. But we'll know the result not in 90 days, but in 90 weeks, when Terafab either works or becomes the most expensive monument to ambition in history.
— Editorial Team
No comments yet.