AI's economics don't make sense
Skepticism is growing over whether current generative AI business models can ever justify the massive capital poured into data centers and model training. Commenters debate if usage-based token pricing, high subscription tiers, and enterprise spend can cover ongoing R&D and infrastructure, or if the industry is effectively subsidizing users while betting that inference costs will keep dropping. Others argue that even expensive AI remains economically viable for many white‑collar tasks, especially coding, but concede that long-term profitability, competitive pressure from open models, and the risk of an eventual correction remain open questions.
Subscription vs Usage-Based Pricing
- Many see flat-fee “all you can eat” plans as fundamentally misaligned with AI’s variable compute costs; heavy users (agents, coding tools) can burn hundreds of dollars a day.
- Others note all subscriptions cross‑subsidize light users to heavy ones; that alone doesn’t make the model broken.
- Several predict a shift to “electricity-style” metering by tokens or tasks, with tiered plans and better controls for overruns.
- A recurring user concern: metered billing makes it hard to predict costs, echoing past surprises with cloud/hosting.
Token Costs, Margins, and Profitability
- One side claims frontier labs enjoy very high gross margins per token; current API prices are far above marginal inference cost, especially with caching and hardware advances.
- Others counter that even if marginal tokens are profitable, huge R&D and datacenter capex leave companies overall cash‑negative; training must be treated like an expensive, ongoing requirement, not a one‑off asset.
- There is disagreement whether claims about “profitable models over their lifecycle” are meaningful without audited financials.
Scale, Capex, and Bubble Risk
- Skeptics argue valuations and build‑outs (tens or hundreds of billions) cannot be paid back with modest per‑seat pricing; back‑of‑envelope math suggests payoffs stretched over decades, if ever.
- Some compare this to WeWork or other overhyped sectors; others to capital‑intensive industries like pharma or semiconductors where big R&D is normal.
- Many expect consolidation; not all current players will survive.
Value to Users and Employers
- Multiple comments say coding assistants and agents already deliver significant productivity gains, especially for well‑paid knowledge workers, making even high hourly AI costs justifiable.
- Others stress diminishing returns: more generated code can overload review and process bottlenecks; junior hires or process fixes may beat more tokens.
- There’s concern that AI may depress wages while core living costs still rise.
Competition: Frontier vs Open / Local
- Several note strong open‑weight models approaching frontier quality but requiring hefty hardware; local or shared clusters may become attractive for companies with predictable heavy use.
- Intermediaries that resell frontier APIs (e.g., dev tools, search wrappers) are seen as especially exposed if they pay full retail for tokens.
Business Models and Monetization
- Some expect advertising or subtle “sponsored” outputs to emerge, analogizing to search and other media. Others note this could sharply erode utility, especially for agentic use.
- A few argue that companies may ultimately aim more at large enterprise and government contracts than consumers.