Anthropic is expanding to Colossus2. Will use GB200

xAI’s Compute, Grok Demand, and Strategy

• Several commenters say xAI massively overbuilt compute (hundreds of thousands of GPUs, low utilization) and overestimated demand for Grok, which is viewed as a second‑tier model with weak traction outside the social platform.
• Others note some users like Grok for being less censored and good for casual Q&A, but agree it is not state of the art, especially for coding.
• Renting out Colossus 1 and 2 to Anthropic and Cursor is framed as monetizing idle capacity and buying time while xAI retrains or “fixes” its models.

Anthropic’s Needs and Competitive Position

• Anthropic is portrayed as compute‑starved but product‑strong, with rapid revenue growth and high demand for Clifford/Claude models.
• Renting instead of building datacenters is seen as de‑risking: avoids owning depreciating hardware and simultaneously deprives xAI of exclusive compute advantage.
• Some see this as evidence Anthropic is now more of a product company than a pure “frontier lab.”

Financial and IPO Optics

• Many view the deal as a way to juice SpaceX/xAI financials before a SpaceX IPO, reclassifying “hardware rental” as high‑multiple “AI revenue.”
• Others push back: revenue is real, disclosed as datacenter rental, and large enough to plausibly pay back tens of billions in capex within a few years, depending on margins.

Ethics, Politics, and Reputation

• Strong criticism that Anthropic is compromising its stated safety/ethics stance by partnering with xAI’s owner, including over:
  • Alignment and political bias of Grok.
  • Legal/ethical controversies around the social platform and broader politics.
• Some argue all major AI founders are similarly power‑ and monopoly‑seeking; none should be idealized.

Environmental and Local Impact

• Heavy focus on Colossus gas‑turbine power plants using permitting loopholes; concerns about NOx, particulates, and noise for nearby residents.
• Some minimize the global impact (“rounding error”); others emphasize cumulative local harms and regulatory failure.
• Debate over whether blame lies more with datacenter operators exploiting loopholes or with slow, constrained grid expansion.

Model Security and Exfiltration Risk

• Users ask whether the datacenter operator could steal Anthropic’s model weights or token streams.
• Responses: weights are said to be encrypted and protected by contracts, but several note that:
  • Inference requires decryption on GPUs, so a determined operator could theoretically exfiltrate with enough effort.
  • Side‑channel or distillation attacks via APIs are likely easier and already alleged elsewhere.
• Consensus: legal, reputational, and business risks make overt theft unlikely, but not technically impossible.

Macro Signals and AI Bubble Concerns

• Some interpret xAI renting out core compute as a bearish signal for Grok and even for AI infrastructure demand forecasting.
• Others argue demand is still massive; the real story is misaligned timing and execution, not lack of long‑term need.