What Colossus 2’s Million-Chip Plan Could Mean for AI in 2027

A building that once held Hormel products could soon house roughly 200,000 AI processors, devoted not to training a new model but to running models for users. That is Elon Musk’s description of “Spam,” one part of the Colossus 2 expansion he outlined on October 31, 2026.

That distinction is the interesting part. The buildout is not only about making the next AI more capable. It could also change how much work existing models can afford to do: exploring several solutions, checking their answers, and staying with a difficult task for longer.

Musk’s broader target is about 1.2 million NVIDIA Blackwell processors at Colossus 2 by the end of 2026. It is a company projection, not an independently verified operating inventory. But it poses a worthwhile question: if that capacity becomes usable, what could it make possible in 2027?

What Musk Is Actually Adding

In his building-by-building update, Musk described four facilities:

• Macroharder: 220,000 GB300s configured for training.
• Minihard: another 220,000 GB300s in the same training configuration.
• Gun Range: a planned 220,000 GB300s for training, targeting operation by the end of November 2026.
• Spam: approximately 200,000 GB300s for inference, targeting December 2026. Inference means running a trained model to produce answers or perform tasks.

That adds up to approximately 860,000 GB300s, including the two planned additions. Musk separately gave the campus-wide target of about 1.2 million Blackwells. The four-building list omits the older Macrohard building; it is not a complete campus inventory, and Blackwell is a broader hardware family than this GB300 subtotal.

Neither number describes one giant model-training run. Even the 660,000 processors designated for training across three buildings could be divided among models, experiments and customers. The announcement does not disclose their allocation or show that all three buildings can train one model together efficiently.

How That Compares With Model Training

The most useful recent comparison comes from xAI itself. Its July 16, 2026 account of Grok 4.5 says the model was trained across tens of thousands of GB300 GPUs. A GPU is the type of processor doing much of this AI computation. One claimed 220,000-processor training building is a substantially larger hardware pool than that disclosed training allocation, using the same named GPU generation.

There is no honest exact multiplier: “tens of thousands” is not a precise number, and training duration matters. The newer Grok 4.7, announced on September 21, uses a larger base model and longer reinforcement-learning training, according to xAI, but its announcement gives no comparable chip count or total training budget.

For more detailed figures, we have to turn to older, unusually transparent examples:

• Meta’s Llama 3.1 405B, released in July 2024: training used up to 16,384 H100 GPUs. Its model card reports 30.84 million H100 GPU-hours for the 405B model. These figures describe a historical training effort, not the strongest model available today. Technical paper, Meta model card.
• DeepSeek-V3, released in December 2024: its authors report 2,048 H800 GPUs and 2.664 million GPU-hours for pretraining, the initial learning of a base model. Their accounting excludes earlier research and experiments used to develop the recipe. Technical report.

A GPU-hour means one GPU working for one hour. It adds a time dimension that a building’s chip count leaves out. But an H800-hour, an H100-hour and a GB300-hour are not interchangeable units of computing work: the processors, software and workloads differ.

The public documents inspected for OpenAI’s Astra, Anthropic’s Fable 5.1 and Google’s Gemini 4 Argon do not disclose comparable total training budgets. Google also uses its own TPU accelerators. We cannot turn the Colossus announcement into a defensible ranking of how many times it exceeds the training of today’s best models.

What we can say is already striking: infrastructure is being planned at a scale far beyond the processor counts of well-documented earlier training runs. The opportunity is larger than simply spending all of it on one successor model.

Three Ways to Turn More Compute Into Better AI

First, develop stronger models and test more ideas. A larger budget can support a more demanding training run. It can also let researchers compare different architectures, data mixtures and learning methods instead of betting everything on one recipe. A laboratory gains both a larger experiment and more chances to discover which experiment is worth running.

This is why processor counts cannot settle the outcome. In its Grok 4.5 account, xAI emphasizes filtering and selecting training data alongside the hardware. More examples help only when they teach something useful. More computation can improve the search for a good training method, not replace that method.

Second, give models more practice with feedback. Reinforcement learning trains a model through attempted tasks and rewards. A coding model can propose a solution, run tests and learn from which attempts succeed. Computation pays for the attempts, the checks and the subsequent learning. Selected model-generated solutions can also become training examples.

DeepSeek’s 2025 Nature paper on R1 demonstrates how powerful this can be in a well-defined setting. During reinforcement learning, R1-Zero’s average single-answer score on the AIME 2024 mathematics test rose from 15.6% to 77.9%. That is a result on a particular benchmark, not a measurement of general intelligence. It nevertheless shows that substantial gains can come from training a model to use what it has learned, rather than only feeding it more text.

The quality of the feedback matters. The same research describes cases where models learned to exploit an imperfect reward system. More practice with a misleading score can reinforce the wrong behavior.

Third, spend more computation while the model is working. Training changes the model. Inference uses it. A trained model can spend more effort exploring an answer, use tools to check a result, or work alongside other model instances on different parts of a task. Its underlying learned parameters need not change for that extra effort to improve the result.

Anthropic’s 2025 account of its multi-agent research system offers a concrete example. A lead model delegated research to other models and performed better than a single-agent setup on the company’s internal research evaluation. The expense was substantial: multi-agent systems typically used about 15 times as many tokens as ordinary chats in its observations. Tokens are the pieces of text models process and generate, so that extra activity consumes resources.

This is one reason the inference building matters. More capacity can make intensive workflows available to more people. It can also simply serve more ordinary requests. Which benefit users receive depends on how the operator chooses to spend the capacity.

The Interesting 2027 Scenario Is AI That Can Keep Working

If the expansion arrives and useful learning methods continue to improve, the plausible change is not just a more impressive answer in a chat window. It is a wider range of tasks worth delegating.

Consider a software assistant asked to fix a difficult bug. Additional resources could let it inspect more of the codebase, test competing explanations, run a broader set of checks and revise its patch before returning it. Some steps could run in parallel; others must wait for earlier results. Anthropic’s research report explicitly cautions that tightly interdependent work, including many coding tasks, is less suited to large teams of agents.

The same distinction applies to research. More computational effort can support more candidate algorithms, simulations and analyses. Google’s May 2026 AlphaEvolve update reports useful changes to genome-analysis software, database storage and computational chemistry tools. These are examples of AI-assisted experimentation producing better methods, not merely more generated text. They do not mean physical experiments or drug development accelerate at the same rate as a piece of software.

An especially consequential possibility is a feedback loop: AI helps improve the tools used to build and run AI, freeing resources for further experiments. Our analysis of AI helping develop its successors examines the evidence for that loop and the difficulty of judging whether it is genuinely improving.

For 2027, three outcomes should remain distinct. Existing capabilities could become more widely available. More effort per task could make current models more useful. Better training and experimentation could produce genuinely stronger models. These can reinforce one another, but a campus expansion does not establish the size or arrival date of any particular breakthrough.

The accessible version matters as much as the best demonstration. A system that can solve a demanding problem only at exceptional expense changes less of everyday work than one that can do it reliably within a normal project budget. More capacity creates room to close that gap. It does not guarantee lower prices or better service: demand, costs and commercial decisions still matter.

The Hardware Has to Become Useful Work

There is a reason Musk mentions networking alongside the chips. During large training jobs, processors repeatedly exchange information. Fast processors that spend too much time waiting for one another deliver less useful computation than their specifications suggest.

Musk says Spam’s chosen inference configuration needs simpler networking and less than half the fiber of the training buildings. That is a claim about this installation, not a general rule that inference is easy to connect. NVIDIA’s inference deployment documentation describes workloads that still require fast transfers between machines. Our explanation of inference hardware follows where those delays arise.

Electricity, cooling, working software and sustained operation complete the picture. NVIDIA’s GB300 rack design is liquid-cooled; installing processors is not the same as having the whole system ready to run. A late-2026 commissioning target also leaves further training, evaluation and deployment decisions between new hardware and a finished product.

The harder bottleneck may be deciding which work is actually good. A passing test can miss a defect. Several agents can repeat the same mistake. Generated training examples need selection and verification. More compute expands the number of attempts; trustworthy feedback determines how much useful progress survives them.

That is the opening Colossus 2 could create in 2027: fewer valuable tasks abandoned because trying, checking and trying again costs too much. The important number will not only be the processors in the buildings. It will be how much dependable work people can finally afford to ask them to do.