Why Everybody Is Losing Money On AI

Cursor, Anthropic, and Weird Channel Economics

• Commenters found it striking that Cursor reportedly passes essentially all its revenue to Anthropic, which is both its core supplier and direct competitor.
• Some see this as unsustainable and question what happens to users if Cursor fails; others assume they will just shift to alternative AI coding tools.
• From Anthropic’s side, selling heavily discounted capacity to a reseller who loses money is also seen as odd but consistent with land-grab strategies.

Training vs. Inference and Real Unit Economics

• Several argue that model inference appears to have decent gross margins (e.g., ~50%), and that losses are driven mainly by huge training and research spend.
• Others counter that you can’t ignore ongoing training, data licensing, salaries, and overhead; treating training as a one-off capex is misleading if the competitive race never stops.
• A recurring point: AI breaks the old “software has near-zero marginal cost” assumption—every query consumes costly compute.

Will Costs Come Down?

• One camp insists cost curves will improve via hardware, architectures, and software optimizations, citing massive historical drops in storage/compute prices and recent per‑token price reductions.
• Skeptics argue the article’s point: costs haven’t fallen fast enough so far, structural constraints (GPUs, power, data centers) are real, and not all tech follows a Moore-like curve.
• There’s disagreement over whether current reasoning/agentic usage patterns are erasing per-token price gains.

Why Keep Losing Money? (VC and Strategy Logic)

• Many say this is normal VC behavior: burn cash now to capture market share in a potentially huge, winner-take-most space; analogous to early Amazon or Google.
• Others object that this only makes sense if AI really is a $10T “golden goose,” which some are beginning to doubt.

Profitability, Pricing, and Competition

• Some argue AI could be profitable today if firms stopped training new models and/or raised prices; competition and expectations, not intrinsic economics, keep margins thin.
• Others respond that pausing training would sacrifice freshness and advantage, and that high compute, hardware, and energy costs limit how far prices can rise before demand drops.

Adoption, Value, and Skepticism

• Mixed experiences: some users feel LLMs deliver huge personal value and would pay much more; others have abandoned them with no noticeable loss in productivity.
• Debate over whether AI usage will become a de facto job requirement, similar to IDEs or smartphones, or remain optional for many “boring” software and business tasks.
• A few worry about long‑term dependence on AI platforms that may later become “enshittified” once pricing power is consolidated.

Historical Analogies and Bubble Talk

• Comparisons range from PCs and smartphones (transformative, compounding value) to Segways, Zeppelins, and dot‑com flops (hyped but limited or mispriced).
• Some expect an AI bubble burst that wipes out weak players while leaving underlying behavioral and technical shifts intact.