Anthropic’s S-1 highlights the potential risks of AI products and that even Anthropic with its market-leading AI model and trillion dollar plus valuation faces the same constraints that other tech entrants have faced for decades.
Google, Apple, Amazon, and Microsoft can protect their market position and stymie the growth of innovative entrants via their control of distribution channels, key inputs, and multiple related products that they can preferentially interoperate, bundle, and cross-subsidize.
Control on distribution
Apple’s and Google’s iron grip on app distribution on smartphones costs consumers billions and hinders innovation. Google’s ability to pay to be the default service on iPhones, carriers, OEMs and browsers, for both general search and AI, is a barrier for rivals to get most consumers to even try their products.
The distribution bottleneck for Anthropic is the cloud and it is dependent on the good graces of Google, Amazon, and Microsoft. For the six months ending in June 2026, 48% of Anthropic’s total revenue and 58% of Anthropic’s net accounts receivables came from Amazon Bedrock, Gemini Enterprise Agent Platform, and Microsoft Foundry (S-1, page 56). Anthropic also pays a distribution fee to each of these companies though the S-1 does not reveal how much. These CSPs (cloud service providers) also get to decide on the relative “…pricing, commercial terms, product placement, integration, or access to compute or sales incentives…” for Anthropic’s models compared to their own models and other companies’ models (S-1, page 73).
Control of key inputs
Google’s unparalleled access to user data is key factor in its ability to maintain its general search monopoly and rival access to this user data is one of the remedies in the Google search case.
The key input bottleneck for a frontier model developer like Anthropic is compute. As Anthropic explains (S-1 page 30):
We compete for access to compute with some of the world’s largest and most well-capitalized technology companies many of which develop their own competitive AI models and have significantly greater financial resources than us. Many own and operate cloud computing data center infrastructure, maintain captive chip manufacturing or design capabilities, have long-standing strategic relationships with key hardware suppliers…
Control of related products
The DOJ’s antitrust lawsuit against Apple illustrates how Apple maintains its market power by degrading interoperability for competing products and services and shutting out innovation and competition from its ecosystem.
Anthropic faces a similar situation (S-1, page 40):
…competitors may bundle AI capabilities with other products, integrate AI functionality into products that customers already use, offer closed technology platforms that lock customers into their ecosystems, sell products at zero or negative margins to gain market share, or use their substantial financial resources to subsidize the cost of AI models, platforms, and applications.
Google integrates Gemini into search, Android, and workspace and Microsoft integrates co-pilot into its office suite of products. Apple’s multiyear multi-billion-dollar deal with Google to incorporate Gemini into Siri softens competition between Apple and Google and locks up important distribution channels (Android and iOS) for AI agents.
Anthropic’s mitigation strategy
Part of Anthropic’s strategy is to vertically integrate and develop its own captive compute via what it calls ‘Anthropic Direct’ (S-1, page 74):
To help reduce our dependence on CSPs, address constraints on available compute, and secure access to compute at the scale necessary for frontier model development, we are pursuing initiatives to develop our own dedicated on-premise data center infrastructure.
This capital-intensive strategy is dependent on the success of Anthropic’s IPO and maintaining off the charts revenue growth (S-1, page 22): $41 million (2023), $4.6 billion (2025) and $16.3 billion (first half of 2026). A risky strategy and one that may still not be enough as Anthropic acknowledges (S-1, page 205):
To secure certain data center development projects, which require significant capital commitments in advance of deployment, we have in some cases relied, and expect to continue to rely, on arrangements with financing sources or partners who guarantee the lease obligations to support the development of data center capacity that is expected to host such partner’s chips or computing infrastructure. Our ability to expand compute and infrastructure capacity through these arrangements in the future may depend on the continued willingness and ability of a limited number of CSPs, partners, and financing sources to enter into and perform under them.
The problem is structural and requires a policy response
If the only path to competing with Big Tech is to replicate their vertically integrated business models and to develop related products like search or office productivity suites or a mobile operating system, then this key technology is likely be controlled by a few firms. We need pro-competitive policies that:
require the largest CSPs to offer non-discriminatory access to compute and cloud (to frontier model developers and companies deploying open-source models);
restrict tying or preferential interoperability of a CSP’s own AI models to its cloud infrastructure; and
open-up distribution channels and stop accommodations among gatekeepers like Apple and Google.
Setting safety standards and public testing capacity is also crucial – not just for safety but also for competition.
First, the S-1 spells out the potential risks of AI products, and it should make clear that while enforcing and clarifying AI product liability is necessary, it will not be sufficient. The S-1 is a siren call to Congress to stand-up and invest in an appropriate safety regime to reduce the probability of catastrophic cyber-attacks, the use of bioweapons or other misuse by bad actors.
Our allies like the UK are already investing in a new National AI Biosecurity initiative, the AI Security Institute, and a National Cyber Shield. We should work with the UK and other close allies to jointly build out such expertise in US agencies like NIST. Building on this safety regime we should work with China to establish international safety protocols for this technology.
Yes, there might be some capture of such government led safety standards by incumbents, but this will be unambiguously better than letting a cabal of companies self-regulate as the current administration proposes.
Second, there is a strong pro-competitive case for setting safety standards and developing public pre-deployment testing capacity which is not the same as mandatory regulation and licensing. This can be pro-competitive by:
enabling entry by letting start-ups, including open-source deployments, signal quality and safety;
helping limit or better define liability; and
allowing sharing of fixed costs, knowledge, and complementing interoperability rules for compute and cloud distribution so that safety standards do not further entrench the large CSPs.
There is work to be done, and urgently, to ensure that the fruits of this transformative technology are not captured by a few firms and individuals. Ensuring that competition thrives, setting up safety standards, and public testing capacity are essential to ensuring we all benefit, and safely, in the productivity gains, economic growth and scientific advances that this break-through technology promises.