Quick pulse for Monday, October 12. Short takes, with credit to where each one came from.

SpaceX is building the biggest neocloud#

InvestAnswers focuses on SpaceX compute, not the phone-carrier angle from yesterday.

He puts live capacity at about 1.4GW at the end of June, with a path past 2GW by year end and toward roughly 10GW by end of 2027. Goldman, in his read, says SpaceX already charges the highest rents for its compute. His take: that fits a fleet that is unusually coherent and homogeneous. (InvestAnswers)

Separately, he builds a Colossus table from an Elon tweet and lands on about 2.7GW by year end if the remaining GB300 blocks come online as sketched. That is his math, not a company filing. (InvestAnswers)

TSMC’s September print keeps the AI trade alive#

TSMC reported September consolidated sales of about NT$511.86 billion, up 54.6% from a year earlier. (TSMC)

InvestAnswers calls the print good but mostly priced in. His take: the AI trade is still alive. (InvestAnswers)

Every built one agent for the whole company#

Every replaced personal Plus One bots with a shared Every Agent that works in public Slack threads. Colleagues see the request and the corrections. It runs on Anthropic Claude Managed Agents (Claude-only; no zero-data-retention option in their write-up). A model upgrade plus small code fixes cut token costs across 11 common tasks by more than 80%. Each teammate gets $15 in free credits, with 0% markup on tokens. (Every)

Haiku 5.5 and the cost-per-task race#

Anthropic launched Claude Haiku 5.5. On OSWorld 2.1 (offline subset) it scores 72.4%, up from 15.7% for Haiku 4.5. For prompts up to 100k tokens, list prices are about 90% below Haiku 4.5. It is also the first Haiku with an adjustable effort setting. (Anthropic)

On the Hermes Index from Nous Research, models are ranked by score and dollars per finished task inside the same agent harness. Claude Opus 5.5 leads at 63.31 for about $4.99 per task; GPT-6 Astra sits at 56.25 for about $11.61; DeepSeek V4.1 Flash lands at 36.91 for about $0.26. (Hermes Index)

Geoffrey Huntley: ship a CLI, then get into the weights#

Geoffrey Huntley sorts developer-tool companies into S-tier and F-tier by whether the model already knows them. GitHub is S-tier because the behavior sits in the weights; a new CLI that needs an MCP or a skill pack starts in F-tier. His playbook: build a strong CLI, tune --help so an agent can walk the verb tree with fewer tool calls, publish docs an agent can finish in one search, then contract with labs for the next training run. (Geoffrey Huntley)

Milk Road: the bear caller is buying again#

Milk Road flags Caleb Franzen flipping back to bull. He sold a large Bitcoin stack near $97K last November, bought between $60K and $75K, and now says a new bull market holds if BTC keeps the 200-day MA cloud around $71K. A limit filled at $83K; his upside sketch runs as high as $205K. Framing: risk about 15% downside for a shot at roughly 130% upside. (Milk Road)

How crowded are US AI data centers, really?#

InvestAnswers counts 139 AI data centers in the US. That is about three per 100,000 square miles, or roughly one per West Virginia-sized patch. His take: public panic about sprawl is overblown. Separately, he flags chatter that Google and other frontier labs are near a recursive self-improvement loop. Insiders claim it; he hedges and still says buckle in. (InvestAnswers, InvestAnswers)

-Q