AI has a subsidization problem

AI has a subsidization problem

High-level thesis

• Cloud inference is not free. Although per-token/model prices have fallen, total cost per request has risen dramatically because prompts, tool calls, and multi-step reasoning generate far more tokens and much longer GPU time.
• Big AI companies subsidized inference to win users (a land grab). Those subsidies are economically unsustainable and are being rolled back — causing capacity limits, feature restrictions, and policy-enforcement changes that affect free and low-cost users first.
• Subsidies are done for three main reasons:
  1. Ad revenue (usually insufficient at scale).
  2. Data collection (valuable but limited).
  3. Stealing customers/market share (the primary motive).

Immediate product / service changes called out

• Gemini (Google)

  • Gemini CLI updates: stronger detection of policy-violating use cases, prioritization of traffic for certain account types, and restriction of Gemini Pro models for free-tier users.
  • Throttling/limits on free/subsidized usage; paid subscribers have also seen capacity problems (e.g., unable to access new Pro models).
  • Official recommendation: use your own paid API key (AI Studio or Vertex.ai) for direct quota and billing control.

• GitHub Copilot / Student plan

  • Complimentary Copilot access moved to a GitHub Copilot Student plan.
  • Some premium models (transcript mentions models like “GPT 5.4”, “Claude Opus”, “Sonnet”) will no longer be selectable under the student plan.

• Open code / Anthropic integrations

  • Open code 1.3.0 will no longer load the “Claude Max” plugin.
  • Anthropic (Claude) resisted allowing developer choice; companies are restricting third-party harnesses/plugins that might enable switching or hybrid use.

• T3 products (host’s own)

  • T3 Chat and T3 Code discussed as paid offerings; example given where per-message billing can be mismatched with per-token API costs (a single heavy prompt can exhaust subscription value).
  • T3 Code integrates multiple backends and provides an open-source wrapper; users are urged to consider paid keys and subscriptions.

Economic and technical analysis (key points)

• Per-token model price has fallen, but token usage per interaction has increased 10x–100x in many real-world uses (code, long documents, tool chains). That can raise cost per interaction despite cheaper token prices.
• Billing model mismatch: per-message pricing (used by some product subscriptions) can be wildly misaligned with per-token API costs — a few heavy prompts can consume most of a subscription.
• Free/subsidized users are typically low-value:
  • Many never convert to paid plans but consume support and GPU capacity.
  • Large free quotas attract many non-paying users and can overload capacity, harming paid customers.
• Ads cannot realistically subsidize heavy inference at the per-request level — ad revenue per view/user is too small to cover multi-dollar requests at scale.
• Data collection is a motive for subsidization but rarely sufficient alone to justify broad free inference.
• The main effective motive for massive free/subsidized allocations is customer acquisition (land grab); this explains temporary heavy subsidization and rapid policy reversals when costs/usage become unwieldy.

Examples and anecdotes

• Paid subscribers (including the host) were reported not getting access to Gemini 3.1 Pro because Google had been giving too much away.
• Anthropic’s $200/month plan can include very large compute allowances (host cites up to ~$5,000 of compute quota) — a massive subsidization intended to lock users in, but risky if users don’t convert to long-term revenue.
• OpenAI temporarily raised rate limits (2x) on some services as a competitive tactic to attract users from rivals.

Practical guidance and recommendations

• For reliable quota and control, use your own paid API keys (AI Studio, Vertex, OpenAI) rather than relying on CLI/embedded subsidized tiers.
• Be cautious with per-message subscription models; heavy, token-dense workloads should use token-based billing or dedicated API keys to avoid surprises.
• If you can afford it and rely on these tools, consider maximizing paid subscriptions while heavy subsidies still exist — but expect those subsidies to disappear.
• Developers/builders: there’s opportunity to build because major providers are reducing subsidies and capacity is under pressure — plan for paid or self-hosted strategies.

Product / tool features called out

• Gemini CLI: policy detection, account traffic prioritization, Pro model restrictions.
• GitHub Copilot Student: new plan with reduced selectable premium models.
• Open code 1.3.0: stops loading Claude Max plugin.
• T3 Chat / T3 Code: open-source wrapper, multi-backend support, per-message vs per-token billing concerns.
• Depotci (sponsor): alternative to GitHub Actions — faster CI, improved Docker builds, CLI and SSH into runners, dashboards, two-command migration pitch.
• Ror Max (sponsor): cloud-assisted iOS app builder, Mac app + phone linking, two-click App Store deployment, supports Swift and AR features.

Takeaways

• The era of broad, deep subsidization of inference is ending or being heavily curtailed: expect more limits, more model restrictions for free users, and more push to paid API keys.
• Businesses will continue to use subsidization selectively for customer acquisition, but long-term sustainability requires converting users and/or dramatically cheaper inference.
• Developers should prepare: prefer paid keys for critical workloads, be wary of free tiers for production, and use remaining subsidies wisely while expecting them to vanish.

Main speakers / sources referenced

• Video host / narrator (channel creator; discusses T3 Chat and T3 Code).
• Google — Gemini, Gemini CLI, Gemini Pro, AI Studio, Vertex.ai.
• Anthropic — Claude (Claude Code, Claude Max, Claude Opus/Sonnet mentioned).
• OpenAI — GPT family, temporary rate-limit changes and partnership incentives.
• GitHub Copilot (student plan / changes).
• T3 (host’s products).
• Depotci and RorMax (sponsors).
• Other referenced players and users: Cursor, “anti-gravity” UI, and example users (Linus Torvalds mentioned).

Note: The transcript was auto-generated; some model and product names may be slightly mis-transcribed.