Welcome to AI Under the Hood — a series where I break down the concepts behind the AI models everyone uses but nobody explains.
Topic 1 → Mixture of Experts
Topic 2 → coming soon
Topic 3 → coming soon
Follow @caprikaps for the series 👇
#AI#GenAI#LLM
@nitinranaa There are tonnes of similar videos talking about dire straits of the game. Just wondering if you’ve also shared those? C26 is clearly a big step up but to older versions which were poor.
@businessbarista Heading AI at a large consulting firm, this is literally our day-to-day with enterprise execs. Org structure, governance, change management, agentic workflows, all of it. Happy to share what we're seeing across clients if useful.
@burkov Or it means Musk is hedging across multiple bets and using compute revenue to fund them. xAI alone wasn't going to compete with Claude/GPT. SpaceX as the AI infrastructure giant has a real shot. He's not giving up on Grok, he's spreading the bet.
@svpino The upfront cost is solved. The hidden cost is what comes back. CLI lets you pipe and filter. MCP returns full structured data every time. Tokens add up fast.
The SaaS economics counterThe math assumes every Pro user is a power user maxing limits. Most aren't. SaaS works because heavy users are minorities subsidized by the majority who barely use the product. Plus enterprise is 80% of Anthropic's $30B ARR. Pro/Max tier is developer pipeline, not profit center.
@elonmusk@SawyerMerritt This is the most vertically integrated AI play on the planet now. Rockets, satellites, data centers, models, products. All one company.
For 6 months, Codex's whole pitch was "use us because Claude rate-limits you." Every marketing thread, every "I switched from Claude" tweet, every Cursor defection, all centered on limits. Today that pitch died in a day. The model war ended a while ago. The real race is compute, and only the compute giants win it.
@sporadica SpaceX needs recurring revenue stories before its fall IPO. Selling Colossus capacity to a major lab is a textbook B2B SaaS revenue line that public markets love. This is probably more about IPO positioning than xAI cap
@doodlestein Even if xAI needs the compute, selling some to Anthropic at premium prices generates revenue that funds even more infrastructure. Microsoft does this with Azure
@KobeissiLetter If governments eventually mandate "AI safety alignment" for infrastructure providers, SpaceX is already positioned. Musk is pre-empting regulation by self-imposing it as a competitive feature. Genius!
@xai@AnthropicAI Selling compute to your direct competitor (Anthropic) before your IPO is a flex move. It signals "we have so much capacity we can power the entire industry." That's a story Wall Street will pay for. My thoughts: https://t.co/6sFv4j2ybK
This isn't just a compute deal. Three things most people will miss:
Anthropic's total committed compute now exceeds 15GW across deals (Amazon 5GW, Google/Broadcom 5GW, Microsoft, now SpaceX). That's industrial-scale infrastructure equivalent to 15 nuclear reactors of dedicated power. The "compute scarcity" era for Anthropic is essentially over.
SpaceX is quietly becoming the AWS of AI before its 2026 IPO. They acquired xAI in January, have a $60B Cursor option, and now serve Anthropic's inference. This is vertical integration nobody else can match: rockets, satellites, AI compute, AI products.
The "orbital AI data center" line was buried in the announcement, but it's the long-term signal. Earth-based data centers are hitting power/cooling/land limits. Starship economics make orbital compute viable within a decade. This deal isn't just about today's GPUs, it's about positioning for the next infrastructure layer.
My take on what this means for intelligence itself:
We've been benchmarking AI progress through model capability (parameters, reasoning, multimodal). That era is ending. The next decade of AI progress will be gated by compute access, not model architecture
When Anthropic, OpenAI, Google, and xAI all have similar capability tiers (and they will, capabilities converge fast in this industry), the differentiator becomes who can serve their model at scale, with low latency, without rationing intelligence. That's an infrastructure problem, not a research one
Today's deal is the moment Anthropic transitioned from supply-constrained to demand-driven. It can finally compete on raw capability without the asterisk of "but you'll hit rate limits." For users, that means Claude becomes more useful per dollar starting now. For the industry, it means compute is the new moat. Models are commoditizing. Infrastructure isn't.
@ClaudeDevs A hug development confirming "Compute" is the next currency as we move to next phase of AI. Have shared my thoughts: https://t.co/6sFv4j2ybK
This isn't just a compute deal. Three things most people will miss:
Anthropic's total committed compute now exceeds 15GW across deals (Amazon 5GW, Google/Broadcom 5GW, Microsoft, now SpaceX). That's industrial-scale infrastructure equivalent to 15 nuclear reactors of dedicated power. The "compute scarcity" era for Anthropic is essentially over.
SpaceX is quietly becoming the AWS of AI before its 2026 IPO. They acquired xAI in January, have a $60B Cursor option, and now serve Anthropic's inference. This is vertical integration nobody else can match: rockets, satellites, AI compute, AI products.
The "orbital AI data center" line was buried in the announcement, but it's the long-term signal. Earth-based data centers are hitting power/cooling/land limits. Starship economics make orbital compute viable within a decade. This deal isn't just about today's GPUs, it's about positioning for the next infrastructure layer.
My take on what this means for intelligence itself:
We've been benchmarking AI progress through model capability (parameters, reasoning, multimodal). That era is ending. The next decade of AI progress will be gated by compute access, not model architecture
When Anthropic, OpenAI, Google, and xAI all have similar capability tiers (and they will, capabilities converge fast in this industry), the differentiator becomes who can serve their model at scale, with low latency, without rationing intelligence. That's an infrastructure problem, not a research one
Today's deal is the moment Anthropic transitioned from supply-constrained to demand-driven. It can finally compete on raw capability without the asterisk of "but you'll hit rate limits." For users, that means Claude becomes more useful per dollar starting now. For the industry, it means compute is the new moat. Models are commoditizing. Infrastructure isn't.
This isn't just a compute deal. Three things most people will miss:
Anthropic's total committed compute now exceeds 15GW across deals (Amazon 5GW, Google/Broadcom 5GW, Microsoft, now SpaceX). That's industrial-scale infrastructure equivalent to 15 nuclear reactors of dedicated power. The "compute scarcity" era for Anthropic is essentially over.
SpaceX is quietly becoming the AWS of AI before its 2026 IPO. They acquired xAI in January, have a $60B Cursor option, and now serve Anthropic's inference. This is vertical integration nobody else can match: rockets, satellites, AI compute, AI products.
The "orbital AI data center" line was buried in the announcement, but it's the long-term signal. Earth-based data centers are hitting power/cooling/land limits. Starship economics make orbital compute viable within a decade. This deal isn't just about today's GPUs, it's about positioning for the next infrastructure layer.
My take on what this means for intelligence itself:
We've been benchmarking AI progress through model capability (parameters, reasoning, multimodal). That era is ending. The next decade of AI progress will be gated by compute access, not model architecture
When Anthropic, OpenAI, Google, and xAI all have similar capability tiers (and they will, capabilities converge fast in this industry), the differentiator becomes who can serve their model at scale, with low latency, without rationing intelligence. That's an infrastructure problem, not a research one
Today's deal is the moment Anthropic transitioned from supply-constrained to demand-driven. It can finally compete on raw capability without the asterisk of "but you'll hit rate limits." For users, that means Claude becomes more useful per dollar starting now. For the industry, it means compute is the new moat. Models are commoditizing. Infrastructure isn't.
This isn't just a compute deal. Three things most people will miss:
Anthropic's total committed compute now exceeds 15GW across deals (Amazon 5GW, Google/Broadcom 5GW, Microsoft, now SpaceX). That's industrial-scale infrastructure equivalent to 15 nuclear reactors of dedicated power. The "compute scarcity" era for Anthropic is essentially over.
SpaceX is quietly becoming the AWS of AI before its 2026 IPO. They acquired xAI in January, have a $60B Cursor option, and now serve Anthropic's inference. This is vertical integration nobody else can match: rockets, satellites, AI compute, AI products.
The "orbital AI data center" line was buried in the announcement, but it's the long-term signal. Earth-based data centers are hitting power/cooling/land limits. Starship economics make orbital compute viable within a decade. This deal isn't just about today's GPUs, it's about positioning for the next infrastructure layer.
My take on what this means for intelligence itself:
We've been benchmarking AI progress through model capability (parameters, reasoning, multimodal). That era is ending. The next decade of AI progress will be gated by compute access, not model architecture
When Anthropic, OpenAI, Google, and xAI all have similar capability tiers (and they will, capabilities converge fast in this industry), the differentiator becomes who can serve their model at scale, with low latency, without rationing intelligence. That's an infrastructure problem, not a research one
Today's deal is the moment Anthropic transitioned from supply-constrained to demand-driven. It can finally compete on raw capability without the asterisk of "but you'll hit rate limits." For users, that means Claude becomes more useful per dollar starting now. For the industry, it means compute is the new moat. Models are commoditizing. Infrastructure isn't.
@claudeai@SpaceX A significant development - have shared my take on this and this could be a turning point as we move to next phase of AI from "Models" to "Compute" https://t.co/6sFv4j2ybK
This isn't just a compute deal. Three things most people will miss:
Anthropic's total committed compute now exceeds 15GW across deals (Amazon 5GW, Google/Broadcom 5GW, Microsoft, now SpaceX). That's industrial-scale infrastructure equivalent to 15 nuclear reactors of dedicated power. The "compute scarcity" era for Anthropic is essentially over.
SpaceX is quietly becoming the AWS of AI before its 2026 IPO. They acquired xAI in January, have a $60B Cursor option, and now serve Anthropic's inference. This is vertical integration nobody else can match: rockets, satellites, AI compute, AI products.
The "orbital AI data center" line was buried in the announcement, but it's the long-term signal. Earth-based data centers are hitting power/cooling/land limits. Starship economics make orbital compute viable within a decade. This deal isn't just about today's GPUs, it's about positioning for the next infrastructure layer.
My take on what this means for intelligence itself:
We've been benchmarking AI progress through model capability (parameters, reasoning, multimodal). That era is ending. The next decade of AI progress will be gated by compute access, not model architecture
When Anthropic, OpenAI, Google, and xAI all have similar capability tiers (and they will, capabilities converge fast in this industry), the differentiator becomes who can serve their model at scale, with low latency, without rationing intelligence. That's an infrastructure problem, not a research one
Today's deal is the moment Anthropic transitioned from supply-constrained to demand-driven. It can finally compete on raw capability without the asterisk of "but you'll hit rate limits." For users, that means Claude becomes more useful per dollar starting now. For the industry, it means compute is the new moat. Models are commoditizing. Infrastructure isn't.
This isn't just a compute deal. Three things most people will miss:
Anthropic's total committed compute now exceeds 15GW across deals (Amazon 5GW, Google/Broadcom 5GW, Microsoft, now SpaceX). That's industrial-scale infrastructure equivalent to 15 nuclear reactors of dedicated power. The "compute scarcity" era for Anthropic is essentially over.
SpaceX is quietly becoming the AWS of AI before its 2026 IPO. They acquired xAI in January, have a $60B Cursor option, and now serve Anthropic's inference. This is vertical integration nobody else can match: rockets, satellites, AI compute, AI products.
The "orbital AI data center" line was buried in the announcement, but it's the long-term signal. Earth-based data centers are hitting power/cooling/land limits. Starship economics make orbital compute viable within a decade. This deal isn't just about today's GPUs, it's about positioning for the next infrastructure layer.
My take on what this means for intelligence itself:
We've been benchmarking AI progress through model capability (parameters, reasoning, multimodal). That era is ending. The next decade of AI progress will be gated by compute access, not model architecture
When Anthropic, OpenAI, Google, and xAI all have similar capability tiers (and they will, capabilities converge fast in this industry), the differentiator becomes who can serve their model at scale, with low latency, without rationing intelligence. That's an infrastructure problem, not a research one
Today's deal is the moment Anthropic transitioned from supply-constrained to demand-driven. It can finally compete on raw capability without the asterisk of "but you'll hit rate limits." For users, that means Claude becomes more useful per dollar starting now. For the industry, it means compute is the new moat. Models are commoditizing. Infrastructure isn't.
We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity.
This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.
Services as a service is the right read. what it actually changes is the consulting industry — accenture and deloitte built whole practices around "we'll deploy AI for you." now the model labs are doing that work themselves. the moat isn't the model anymore, it's owning the deployment relationship end to end.
Yeah this is genuinely strange. the honest answer might just be that long, coherent, reasoning-heavy documents are rare in training data. you can't teach a model to reason across 500k tokens if your training set is mostly stitched-together unrelated stuff. retrieval got solved early, reasoning over long context didn't, and people kept conflating the two.