Elon Musk’s one-word posts on X this week, “No more AI,” then “SI,” then “It’s better,” set off a scramble to figure out what SpaceX’s AI operation is actually becoming. When a follower asked directly whether SpaceXAI would turn into SpaceXSI, Musk answered, “Yes, we will make that change.” SpaceX itself has not issued a formal statement confirming the rename, and Musk gave no timeline for when it would take effect.
The naming fight is the visible part of the story. The less visible part, and arguably the more consequential one, is what sits behind it: a GPU and data center buildout that has quietly grown into one of the largest private computing projects in the United States. According to research pulled from multiple outlets this week, xAI’s Colossus cluster in Memphis, Tennessee, is on pace to house roughly 1.44 million Nvidia GPUs by late December 2026, backed by tens of billions of dollars in financing and a parallel push to design custom silicon under a project Musk has called Terafab. That is the part of “SpaceX is changing” that actually moves markets, chip orders, and power grids. This piece digs into the hardware reality behind the rebrand chatter: the chips, the money, the memory shortage nobody saw coming, and what it means for Nvidia, Tesla, and everyone else racing to build AI infrastructure at the same time.
What’s Actually Changing at SpaceX’s AI Operation
Start with what is confirmed. Musk’s “No more AI… SI… It’s better” sequence of posts, and his follow-up “Yes, we will make that change” when pressed on SpaceXAI becoming SpaceXSI, are real and verifiable on X. Beyond that, the public record thins out fast. SpaceX has not confirmed the rename in a press release, an SEC-adjacent filing, or any corporate communication. There is no confirmed date for when the SpaceXAI account or branding would shift to SpaceXSI, and claims that SpaceX has completed an acquisition of xAI remain unconfirmed by a primary source, even though several secondary reports describe the two companies as increasingly intertwined.
That ambiguity matters for a simple reason: a name change costs nothing. A 1.44-million-GPU data center cluster costs tens of billions of dollars, draws gigawatts of power, and takes years to plan. If SpaceX and xAI really are consolidating their AI operations under one banner, the hardware side of that consolidation, not the letters in the name, is where the real story lives. For context on the rename itself and how markets reacted to it, see our earlier coverage of SpaceX’s SpaceXSI rename and the stock move that followed.
Inside Colossus: Where SpaceX’s AI Hardware Actually Lives
Colossus is the name attached to a cluster of data centers in the Memphis, Tennessee and Southaven, Mississippi area, built originally under the xAI banner. Public descriptions, including an October inventory cited in industry research, put Colossus 1 at roughly 230,000 GPUs: about 150,000 Nvidia H100 chips, 50,000 H200 chips, and 30,000 GB200 systems. The servers reportedly run on Supermicro liquid-cooled 4U chassis with eight-GPU HGX boards, a design chosen for density and for how fast it can be racked and powered on.
Colossus 2 is the newer, bigger phase, and it is where the Nvidia Blackwell generation comes in. Reports describe Musk characterizing the second cluster as containing 110,000 GB200 chips and 440,000 GB300 chips, with further GB300 deployments arriving in waves of 220,000 units at a time. A Scientific American report from September described Colossus 1 as running more than 220,000 Nvidia GPUs, with Colossus 2 expected to roughly double that footprint once fully built out. None of these figures come from an audited corporate disclosure. They come from Musk’s own statements, trade press, and aggregator sites tracking the buildout in near-real time, so they are best treated as directionally accurate rather than exact to the chip.
From Hopper to Blackwell: The GPU Upgrade Driving Costs
The generational jump matters more than the raw chip count. Nvidia’s H100 and H200 chips belong to the Hopper generation, chips that have powered most large AI training runs since 2023. GB200 and GB300 belong to the newer Blackwell generation, which pairs GPUs with Grace CPUs on a single rack-scale platform connected by high-bandwidth interconnects. Blackwell systems are faster per chip, but they are also hungrier: they need more advanced memory, denser liquid cooling, and heavier-duty networking than the Hopper gear they are replacing.
That tradeoff is exactly what is playing out at Colossus. Reports describe another 660,000 GB300 GPUs planned in three separate 220,000-unit tranches, which, stacked on top of the existing footprint, is how outlets arrive at the roughly 1.44 million GPU figure for the Memphis-area site by late December 2026. If that pace holds, Colossus would be sitting on a fleet of AI accelerators larger than many national research computing programs, built by one group of affiliated companies in roughly two years. Nvidia’s own chip pricing has been moving in the same direction, a trend we covered in our report on Nvidia B200 prices doubling as Tencent and Oracle signed a separate $7 billion deal.
Colossus GPU Buildout: Phase-by-Phase Numbers
The table below lays out the reported progression of the Colossus cluster, from its original Hopper-based footprint through the Blackwell expansion. All figures are as reported by trade press and research aggregators tracking the project. None are confirmed by an audited SpaceX or xAI disclosure.
| Phase | GPU Architecture | Reported GPU Count | Reported Power Draw | Target Timeline |
|---|---|---|---|---|
| Colossus 1 | Nvidia Hopper (H100, H200) + early GB200 | ~230,000 (150K H100, 50K H200, 30K GB200) | ~300 MW energized | Operational, 2024-2026 |
| Colossus 2, initial build | Nvidia Blackwell (GB200, GB300) | 110,000 GB200 + 440,000 GB300 | Part of combined ~1.4 GW footprint | Reported operational target, late 2026 |
| Colossus 2, planned tranches | Nvidia Blackwell (GB300) | +660,000 across three 220,000-unit waves | Additional 500 MW/month of chips cited | Through December 2026 |
| Combined Memphis-area total | Mixed Hopper + Blackwell | ~1.44 million GPUs (reported) | Musk has cited a 10 GW target by end of 2027 | Late December 2026 and beyond |
The power figures are arguably the more telling data point. A combined draw approaching 1.4 gigawatts is not a server room problem, it is a utility-scale infrastructure problem, which is why reports describe the project pairing GPU procurement with turbines, substations, and liquid cooling built on a four-month construction timeline per facility. Our earlier coverage of liquid cooling vendors Vertiv, CoolIT, and Motivair competing for AI data center contracts gives a sense of how much of this spend never shows up in a GPU price tag at all.
The $40 Billion Question: How SpaceX Plans to Pay for the Chips
Buying GPUs at this scale requires financing structures that look more like infrastructure debt than a corporate IT budget. An October report said SpaceX was in discussions to raise as much as $40 billion in debt financing, with Apollo Global Management and a group of major banks reportedly structuring the deal to fund Nvidia hardware purchases for the Colossus expansion. A separate report described a procurement agreement covering roughly 325,000 GPUs that could exceed $40 billion in total value running through May 2029.
On the equity side, xAI reportedly raised $20 billion in August 2026, financing described as part of the capital base supporting continued compute expansion alongside the debt talks. Taken together, these figures sketch a company funding its AI hardware ambitions through a mix of equity raises, project-style debt, and, if the reported merger discussions move forward, closer financial ties to SpaceX and Tesla. We tracked the broader consolidation story in our report on talks of a Tesla-SpaceX merger potentially landing before a December 6 annual meeting, which lines up with the same window Musk has used for Colossus 2 milestones.
Memory, Not GPUs, Is the New Bottleneck
For the past two years, the AI hardware story has mostly been about GPU scarcity. That is starting to change. A report on Musk’s second-quarter SpaceX earnings call said he identified memory supply, not GPU availability or electricity, as the limiting factor constraining further AI expansion. That is a meaningful shift in framing, because Blackwell-generation systems like GB200 and GB300 lean heavily on high-bandwidth memory, and that memory is now in short supply across the entire industry, not just at Colossus.
The timing lines up with a broader memory crunch we have covered separately: HBM4 prices have jumped as AMD, Samsung, and SK Hynix all compete for the same constrained supply, detailed in our report on the HBM4 shortage pushing prices up 117 percent. When one of the largest single GPU buyers in the world says memory, not silicon, is now the constraint, it is a signal that the entire AI buildout, not just Musk’s companies, is running into the same wall at roughly the same time.
Power Before Silicon: The Infrastructure Nobody Sees
Every GPU in Colossus needs electricity before it needs a training job. Reports from August put the combined rated power draw for the SpaceX- and xAI-linked data centers in Memphis and Southaven at around 1.4 gigawatts, with Musk reportedly targeting 10 gigawatts of capacity across his AI operations by the end of 2027. To hit numbers like that, the deployment model described in trade reports emphasizes speed and repeatability: standardized 450-megawatt facilities assembled in roughly four months each, with as many as five large sites planned by the end of 2026.
That speed has a cost. A September Scientific American report described a growing backlash from residents near the Colossus site over air quality, water use, and grid strain, a pattern that echoes what we reported in New Jersey, where a data center operator was sued again over noise and other impacts affecting more than 1,000 nearby homes, detailed in our coverage of the DataOne Vineland data center lawsuit. The pattern is becoming familiar across the industry: AI compute scales faster than the permitting, grid capacity, and community relations needed to support it.
Terafab: Musk’s Bet on Custom Silicon
Nvidia GPUs are not the only hardware path Musk is reportedly pursuing. A Yahoo Finance report said Musk wants Tesla and SpaceX to test 10 chip designs a year under a project described as Terafab, intended to eventually supply chips for Tesla vehicles, Optimus robots, and what the report called AI data centers in space. That phrasing points toward custom silicon optimized for specific, recurring workloads, things like autonomous driving inference, satellite networking, and onboard compression, rather than the general-purpose training work Nvidia GPUs handle at Colossus.
It is worth separating ambition from evidence here. The Terafab reporting describes a strategic goal, not proof that a production chip fabrication facility is already running. Tesla’s existing Dojo program, built around custom training silicon, is a useful comparison point: available reporting does not establish that Colossus is being converted to Dojo hardware, and current Colossus reporting consistently points to Nvidia H100, H200, GB200, and GB300 chips as the operating hardware today. Custom silicon, if it materializes, would likely run alongside Nvidia systems rather than replace them in the near term.
Financing and Capital Behind the SpaceX AI Hardware Push
| Initiative | Reported Scale | Stated Purpose | Status as of Oct. 17, 2026 |
|---|---|---|---|
| xAI equity raise | $20 billion (August 2026) | Fund continued compute expansion | Reportedly completed |
| SpaceX debt financing talks | Up to $40 billion, with Apollo Global Management and major banks | Procure Nvidia hardware for Colossus | Reported in talks, October 2026 |
| Separate GPU procurement deal | ~325,000 GPUs, potentially exceeding $40 billion | Multi-year Nvidia chip supply | Reported, running through May 2029 |
| Terafab custom silicon program | Target of 10 chip designs tested per year | Chips for Tesla vehicles, Optimus robots, space-based AI compute | Reported strategic goal, no confirmed production fab |
Read across the table, the pattern is clear: nearly every dollar is tied to either buying more Nvidia hardware or building the option to eventually need less of it. That duality is not unique to Musk’s companies. It mirrors the strategy Google, Amazon, and Microsoft have each pursued with their own custom silicon programs, built specifically to reduce dependence on a single GPU supplier even while buying that supplier’s chips at record volume.
How SpaceX’s Chip Strategy Compares to Rivals
Every major AI infrastructure player is running some version of the same playbook: buy Nvidia at scale, build custom silicon on the side, and fight for power and memory supply along the way. What sets the SpaceX and xAI approach apart is speed and concentration. Rather than spreading GPU capacity across multiple regions the way Microsoft, Amazon, and Google tend to, Colossus has concentrated an enormous share of its build in a single metro area, Memphis and Southaven, which maximizes construction speed but also concentrates grid, water, and community risk in one place.
Nvidia’s broader chip pricing and allocation decisions sit upstream of all of this, which is part of why Nvidia’s market position outside the US has become its own story. Our report on Nvidia’s China AI chip market share dropping to 34 percent as Huawei climbs to 28 percent shows how export controls and local competition are reshaping where Nvidia’s Blackwell-generation supply actually goes, a factor that indirectly affects how much capacity is available for buildouts like Colossus. Industry coordination efforts, like the one described in our coverage of Nvidia, AMD, Google, and Meta agreeing on a shared data center specification, suggest the entire sector recognizes that no single company, including Musk’s, can solve the power and interconnect problem alone.
Starlink’s Hardware Edge Versus Kuiper and OneWeb
SpaceX’s AI compute buildout is getting the headlines, but its satellite hardware business is where the company has the longest track record of vertical integration. SpaceX designs the Starlink user terminal, the electronically steered antenna, the modem, and the networking stack as one integrated platform, which has let it scale manufacturing and push software-controlled beam steering faster than most competitors. Amazon’s Kuiper system uses a broadly similar approach, with electronically steered flat-panel antennas designed for cost-efficient mass production, making it the closest architectural rival to Starlink on the terminal side.
OneWeb has taken a different path, leaning toward enterprise, government, maritime, and aviation deployments through partner terminals rather than Starlink’s much larger direct-to-consumer installed base. What the available reporting does not support, at least not yet, is a detailed chip-level teardown comparison between the three systems. Claims about which company has the more advanced onboard silicon in its terminals would be speculation rather than fact at this point, so any such claim should be treated with caution regardless of where it is published.
Historical Context: How SpaceX Got Here
SpaceX built its reputation on hardware reuse: landing and reflying Falcon 9 boosters to cut launch costs, then applying the same iterate-fast approach to Starship and Starlink. xAI’s Colossus cluster, first built in 2024 and expanded continuously through 2025 and 2026, applied a similar philosophy to data centers, reportedly standing up its first phase in around 122 days starting with roughly 100,000 Nvidia H100 GPUs, a construction pace that caught much of the data center industry off guard.
What has changed since then is scale and ambition. A cluster that started as a fast-moving AI lab project now sits at the center of a reported multibillion-dollar financing push, a possible corporate rename, and talk of merging with two of Musk’s other companies. Whether or not SpaceXAI formally becomes SpaceXSI on paper, the Memphis-area hardware buildout is the part of this story that is already reshaping Nvidia’s order book, regional power grids, and the memory supply chain.
Market Impact: What This Means for Nvidia, Tesla, and Investors
For Nvidia, Colossus represents one of the largest single-customer commitments to the Blackwell generation reported anywhere in the industry, and a potential $40 billion-plus procurement relationship running through 2029 is the kind of order that shapes manufacturing priorities at the chip level. For Tesla, a reported push toward custom silicon under Terafab, paired with ongoing merger discussions, suggests Musk is trying to build a hardware supply chain that spans cars, robots, satellites, and AI training under one roof, reducing reliance on any single outside vendor over time.
For investors, the signal to watch is not the SpaceXSI name change itself but whether the reported debt financing closes, whether the 1.44 million GPU target for December 2026 is hit on schedule, and whether the memory bottleneck Musk flagged on the earnings call eases or worsens. Any of those three outcomes would move faster through markets than a rebranded account name ever could.
Five Predictions for the Next Six Months
- A formal confirmation of the SpaceXAI-to-SpaceXSI rename is more likely to arrive alongside a hardware milestone, such as Colossus 2 reaching full operational status, than as a standalone announcement.
- Nvidia’s GB300 shipment schedule is the single biggest risk to the reported 1.44 million GPU target. Any slip in Blackwell supply would likely push that milestone into early 2027.
- Memory suppliers tied to HBM4 production, including Samsung and SK Hynix, will gain pricing leverage as Colossus-scale buyers compete with other hyperscalers for the same constrained wafers.
- At least one of the five large facilities planned for 2026 will likely face a delay tied to community pushback or grid interconnection, following the pattern already seen in Memphis and in New Jersey’s DataOne Vineland dispute.
- If the reported Tesla-SpaceX merger discussions progress toward the December 6 annual meeting window, expect a clearer statement on the Terafab custom silicon timeline to accompany it, rather than arriving separately.
What to Watch Next
Three dates and data points matter more than the SpaceXSI name itself over the next few months: whether the reported $40 billion debt financing with Apollo Global Management and partner banks actually closes, whether Colossus 2’s Blackwell buildout stays on pace for its reported late-2026 target, and whether Musk’s 10-gigawatt-by-2027 power goal looks remotely achievable once grid interconnection queues and local permitting get involved. None of those are guaranteed, and all three depend on supply chains, regulators, and financing partners that operate on timelines Musk does not fully control.
Frequently Asked Questions
Is SpaceX actually renaming its AI unit to SpaceXSI?
Musk confirmed in a post on X that the rename will happen, writing “Yes, we will make that change” when asked directly. SpaceX has not issued a formal corporate announcement or given a timeline, so the change is confirmed in intent but not yet in execution.
What is Colossus and why does it matter more than the name change?
Colossus is the GPU data center cluster in Memphis, Tennessee, originally built by xAI, now reportedly housing hundreds of thousands of Nvidia chips with plans to reach roughly 1.44 million GPUs by late December 2026. It represents tens of billions of dollars in real infrastructure spending, which is a far bigger story than what the AI unit is called.
How many GPUs does Colossus actually have right now?
Reports describe Colossus 1 at roughly 230,000 GPUs (a mix of H100, H200, and early GB200 chips), with Colossus 2 adding 110,000 GB200 and 440,000 GB300 chips, plus further planned additions. These figures come from trade press and research trackers, not an audited company disclosure.
Why is memory, not GPU supply, suddenly the bottleneck?
A report on Musk’s second-quarter SpaceX earnings call said he identified memory supply as the limiting factor for AI expansion, ahead of GPUs or electricity. Blackwell-generation chips require large amounts of high-bandwidth memory, which is now constrained across the entire chip industry, not just at Colossus.
What is SpaceX’s Terafab project?
Terafab is a reported effort to have Tesla and SpaceX test 10 custom chip designs a year, aimed at supplying Tesla vehicles, Optimus robots, and AI data centers in space. It reflects a strategic ambition reported by Yahoo Finance, not confirmed proof of an operating production fab.
Has SpaceX actually acquired xAI?
That claim remains unconfirmed by a primary source. Several reports describe the two companies as closely linked and increasingly consolidated, but no formal, confirmed acquisition announcement from SpaceX or xAI exists as of October 17, 2026.
How does this compare to what Google, Amazon, and Microsoft are doing?
All three hyperscalers run a similar dual-track strategy: buying Nvidia GPUs at massive scale while developing custom silicon to reduce long-term dependence on any single supplier. SpaceX and xAI’s version is distinguished mainly by how concentrated and fast-moving the Memphis-area buildout has been compared to the more geographically distributed approach of its rivals.
What could slow down or derail the Colossus buildout?
The most likely risks are a slip in Nvidia’s GB300 Blackwell shipment schedule, continued memory supply constraints tied to HBM4 production, delays in the reported $40 billion financing closing, and local community or grid-interconnection pushback similar to what has already surfaced around the Memphis site.