RBB Says GenAI Competition Needs Price Evidence, Not Just Frontier Control
Key Takeaways
- Assess GenAI competition by capability adjusted prices, product quality, switching, and model choice, not frontier control alone.
- Treat CCIA's evidence as commissioned evidence: useful, but verify price and usage terms in procurement.
- Watch foreclosure theories around inputs, distribution channels, and partnerships, because those are where enforcement can bite.
Why it matters
- ProductProduct leaders should design for model switching and measurable quality gains, not just attach to the current leading model.
- InvestorsInvestors should compare market power claims against evidence on pricing, churn, buyer multihoming, and application revenue.
The US report for CCIA argues regulators should test outcomes such as cheaper capability, better products, and customer choice.
The awkward thing about regulating a fast AI market is that a screenshot can go stale before the lawyer finishes the slide. A regulator can point to a few large labs, cloud providers, and distribution channels. A buyer can point to falling inference costs, leaderboard churn, and a procurement spreadsheet full of model options. Both can be true, which is why the new RBB Economics report matters.
RBB Economics published Generative AI: an assessment of competitive dynamics, US edition, on November 12, 2026, prepared for the Computer & Communications Industry Association. The report does not ask policymakers to stop looking at concentration. It asks them to stop treating concentration as the whole case.
The outcome test RBB wants regulators to use
RBB Economics says its US focused report is the third in a series prepared for CCIA and focuses on the US because that is where most frontier development is happening. The core move is methodological: judge GenAI competition by market outcomes, not just by who owns a frontier model or a distribution channel. RBB says regulators should look for evidence such as “falling prices for a given level of capability,” alongside product improvement, breadth of choice, and ongoing innovation.
That is a polite way of telling antitrust agencies not to confuse a market map with a market diagnosis. RBB examines two layers, AI applications and foundation models, and it separately reviews concerns about input foreclosure, distribution channel foreclosure, and partnerships. Its most quotable line is also its most legally conventional one: “high shares of supply are not in themselves evidence of a competition problem.” High shares can be the start of the file, not the end of the analysis.
RBB also attaches a warning label to its own conclusion. The report is a snapshot based on data available up to early October 2026, in what it calls an unusually fast moving sector. Translation for builders: do not turn this into a permanent immunity badge for your pitch deck. A snapshot is not a safe harbor.
CCIA's evidence is prices, churn, and switching
CCIA, which commissioned the RBB report, says the evidence through October 2026 is consistent with effective competition and supports vigilance rather than broad preemptive regulation. That is not neutral scripture, since CCIA is an industry association. It is still useful evidence if readers keep the label on the jar.
The price evidence is the cleanest part of the case. CCIA says Stanford's AI Index recorded a more than 280 fold decline in the cost of GPT-3.5 level performance in two years, which it describes as a more than 99% decline. CCIA also says GPT-4 level performance fell from about $20 per million tokens in late 2022 to roughly $0.40 by the end of 2025, a 98% drop.
The churn evidence points in the same direction. CCIA says the top position on a widely followed coding leaderboard changed hands 13 times in 22 months, and that more than 90 models from 28 different firms launched in 2025 alone. It also says startups captured about 63% of AI application revenue in 2025, up from 36% a year earlier. On buyer behavior, CCIA says 78% of companies use two or more model families, and 59% use three or more.
Where regulators will still look
The Federal Trade Commission's policy statements page lists a Joint Statement on Competition in Generative AI Foundation Models and AI Products, which is a useful reminder that agencies are not staring only at benchmark charts. RBB's own framework acknowledges the live theories regulators care about: foreclosure of key inputs, foreclosure through distribution channels, and the competitive effects of partnerships. That is where the compliance work starts to look less like economics seminar material and more like contract review.
For AI builders, the practical question is not whether a market is officially competitive in the abstract. It is whether your customers can switch models without rewriting the product, whether your distribution deal blocks rivals in practice, and whether your dependency on a single supplier becomes a business risk your sales team has to explain. If your vendor contract quietly makes model portability fictional, the fact that 28 firms launched models in 2025 will not save you in a procurement review.
For policymakers, the tension is familiar. A market can show falling prices and improving products while still developing chokepoints that matter later. RBB's report argues for outcome based testing, but it does not erase the need to inspect the pipes.
What buyers and builders should actually measure
CIO separately cited Stanford's 2025 AI Index for the same more than 280 fold decline in GPT-3.5 level inference cost between November 2022 and October 2024. Stanford HAI says the AI Index exists to provide unbiased, rigorously vetted, globally sourced data for policymakers, researchers, journalists, executives, and the public. That independent measurement matters because every vendor price page is also a sales document. I have yet to meet a pricing page that volunteers all the unpleasant edge cases.
The useful takeaway is to build a competition dashboard that looks like procurement reality. Track capability adjusted price, not just headline token price. Track output quality and latency for your actual use case, not someone else's leaderboard. Track model family diversity, switching costs, usage limits, and contract terms that affect whether choice is real or decorative.
The next fight will not be whether GenAI is concentrated or competitive as a slogan. It will be whether prices, quality, switching, and access keep improving once the current investment cycle meets ordinary enterprise budgets. Watch the first enforcement theories around foreclosure and partnerships, not the press releases saying everyone welcomes the discussion. Those are usually written by people whose lawyers have already opened a new folder.
Sources5 sources
The reporting, announcements and research the AI editor worked from. Links open the original publisher.
- Generative AI: an assessment of competitive dynamics (US edition)rbbecon.com
- New CCIA Study Finds Generative AI Market Remains Dynamic and Competitive - CCIAccianet.org
- Policy Statementsftc.gov
- Grok Bot's usage limits: Why the fine print matters more than the pricecio.com
- AI Index | Stanford HAIaiindex.stanford.edu
