OpenAI, NVIDIA And Anthropic Just Split. Here's How I'd Spend $20, $60 Or $200.

Overview

The video argues that recent AI moves—OpenAI’s first custom chip (“Habanero”), disputes over developer coding tools (especially Cursor), and Jensen Huang’s response at Nvidia—signal the market splitting into three practical “camps” competing for the user’s workflow, costs, and long-term dependency.

1) OpenAI’s “Habanero” chip: vertical integration (and cost control)

• OpenAI positions Habanero as its first chip for running AI models (inference).
• The creator highlights OpenAI’s claim that it beat Nvidia systems (GB200/GB300 and “Vera Rubin”) on latency/throughput per watt using open-weight model tests.
• A key theme is speed of development and software enablement:
  • Habanero was taped out after ~9 months.
  • OpenAI claims AI-written code (Codex/GPT-type models) generated optimized kernel/workload code 1.5–1.8× faster than human-written versions for some parts.
• Limits remain:
  • Habanero is not replacing Nvidia for training.
  • OpenAI is still a major Nvidia customer (Nvidia claims ~12 GW of planned/committed Nvidia systems through 2030).

Main takeaway: OpenAI is building enough of its own inference stack to improve unit economics, but it still relies on Nvidia for large portions of compute.

2) OpenAI vs Cursor: lock-in risk when ownership changes

• The creator frames OpenAI’s decision to stop providing future models to Cursor (after SpaceX bought Cursor on Nov 12) as a warning about provider rivalry and sudden loss of model availability inside tools developers rely on.
• The argument generalizes: if model access can be removed due to business/ownership shifts, users may become “trapped” inside a tool’s ecosystem.
• Anthropic is cited as a similar example—withdrawing support for “Windsurf”—reinforcing the idea that providers may prioritize their own surfaces over continuity for third-party tool users.

3) Nvidia as the “sell to everyone” core infrastructure

From Jensen Huang’s earnings-call messaging, Nvidia’s strategy is interpreted as:

• not primarily winning every custom-chip fight,
• but being embedded across many clouds, many workloads, and multiple stages of model development.

Nvidia sells:

• training systems
• systems that help improve models
• networking and data-center-scale infrastructure
• and is distributed through major cloud providers.

Therefore:

• even if OpenAI shifts some inference to custom silicon,
• training and other unpredictable/less planned workloads still require Nvidia-style general systems.

Nvidia also benefits from workload churn:

• new model families and shifting between training vs. inference keep general-purpose infrastructure valuable.

4) Anthropic’s “middle camp”: multi-sourcing to manage compute constraints

Anthropic is portrayed as less tightly vertically integrated and more dependent on maintaining flexible supply, including:

• Amazon Trainium
• Google TPU capacity via a Broadcom-built multi-gigawatt agreement
• Microsoft-provided Nvidia capacity
• SpaceX’s Colossus 1 data center (220,000+ Nvidia GPUs) for Claude

The creator claims this reduces dependency risk:

• if one supplier can’t meet needs or prices,
• Anthropic can route work elsewhere.

Cursor blurs camp boundaries:

• Cursor (now under SpaceX AI) helps distribute access to Claude,
• but Cursor also uses other model options (e.g., Grok/Composer) inside the same product ecosystem.

5) The creator’s consumer strategy: fund models, but don’t let them own your memory/files

A major recommendation is how viewers should spend across AI subscriptions:

Core principle: prevent any single AI company from holding the only copy of your memory, files, and instructions.

Suggested approach (“Open Brand” / local or user-controlled memory):

• keep documents/code in user-controlled storage (files/repos),
• keep important instructions in a portable format,
• let AI models read/manage that controlled memory rather than storing your “life” inside one vendor’s app.

Openrouter (bought by Stripe) is mentioned as an option to route requests across multiple models, but the creator emphasizes you should not outsource the entire computing/memory layer.

6) How much to pay ($20 / $60 / $200+) and for what

• ~$20/month
  • pick one primary provider for weekly work
  • keep at least one rival option via a free account
  • don’t brute-force many models for tiny gains
• ~$60/month
  • direct access to OpenAI (~$20)
  • Claude (~$20)
  • Cursor Pro if coding daily
• Over $200/month (creator’s case)
  • buy multiple frontier plans (e.g., Anthropic, Codex, Grok)
  • expect each to pay back through productivity/time savings; if not, reduce over time
  • the creator mentions OpenAI enterprise moving toward outcome-based pricing (pay only when outcomes are achieved), aligned with a “pay for value” threshold

Overall conclusion

The video’s “three camps” framework is:

1. OpenAI: build more of the full inference stack and control availability on its surfaces to improve economics.
2. Nvidia: remain the interoperable, widely sold infrastructure layer that rival ecosystems still need.
3. Anthropic: maintain flexibility via multi-sourcing to switch compute capacity and avoid being captive to one provider.

Practical consumer message: spend for access to high-quality models, but architect your workflow so model/provider changes don’t break your ability to work.

Presenters / contributors

• Nate B Jones (host/creator; the speaker throughout the subtitles)
• Jensen Huang (referenced)
• Hal Tan (referenced)
• Anthropic / Claude team (referenced)
• SpaceX AI (referenced)
• OpenAI (referenced)
• Nvidia (referenced)
• Elon Musk (referenced as having bought Cursor)
• Openrouter / Stripe (referenced)