AI Foundation Models: A Roundup on the Evolving Competition Landscape

Dr. (Prof.) T.S. Somashekar is the Director, Centre for Competition and Regulation, NLSIU

I.              Background

An inescapable part of our lives is the way AI influences daily activities, directly or indirectly, in incredibly diverse ways, with enormous implications for productivity across sectors and for direct consumer benefit. The pace of technological evolution is rapid in this segment, and tech firms are quickly adjusting their strategies to retain or enhance their market power. Regulators, in turn, have approached these developments from both an ex ante and ex post perspective. Horizontal ‘relationships’/dependencies and vertical integration in the AI ecosystem can increase the potential for downstream anticompetitive effects. Such ‘relationships’ would traditionally be investigated through merger dynamics testing to assess potential unilateral and coordinated effects that enhance market power. Awake to this fact, big tech companies, predominantly Hyperscale’s (massive cloud computing firms with large networks of data centres), with their large proprietary data and compute resources, resorted to novel methods of horizontal ‘partnerships’ with Foundation Models (FMs), termed ‘reverse acquihires’.[1]

FMs are algorithms trained in massive and diverse unstructured data sets and serve as the ‘foundation’ for multiple tasks, unlike traditional ‘Narrow AI’ which are specialised, focusing on individual tasks such as credit card fraud detection or a stock price predictor and so on. By nature, they require massive capex funding for compute power, data to train and deploy models and other infrastructure, including cooling. This creates bottlenecks, as very few firms are capable of providing FM developers with the funding and complementary inputs they need. Proprietary data, so essential for training and deployment/inference, is dominated by the large tech Hyperscale’s, such as Microsoft Azure, Amazon Web Services and Google Cloud Computing,   the compute/accelerator layer is dominated by Nvidia’s GPUs with a few fringe players (AMD’s Instinct series) and relatively a few more players in the ASICs segment (used for the specialized inference stage) such as Google TPU — and distribution, through ‘operating systems’ and office suites, again by the same firms that dominate proprietary data. The fabrication of chips/semiconductors itself is concentrated with TSMC in Taiwan, with concerns of high geopolitical risk.

Recognising the need to be fleet-footed, regulators across countries have undertaken studies and investigations related to potential horizontal (cartelisation – hub and spoke coordination), vertical (exclusionary) anticompetitive issues, relationship/agreements between complementary input providers, hyperscalers and FMs, and potential ‘killer acquisition’ situations.[2] The CMA paper highlighted certain potential concerns arising out of the agreements between Hyperscalers and FMs:  restriction of choice of FMs in downstream markets; restricting access to FMs to prevent competition for themselves and entrenching the position of existing players. Similar concerns have been raised by the US FTC, i.e., restriction of access to inputs to non-partner AI developers; creation of switching costs for the partner FM and AI firms, etc. The Microsoft/Inflection agreement, an example of ‘reverse acquihire’ and ‘killer acquisition ’, was recognised as a merger but was cleared by the UK CMA and the EU as it did not have potential unilateral effects. A similar investigation into the agreement between Alphabet Inc and Anthropic PBC, which involved the acquisition of non-voting shares of Alphabet and non-exclusive compute and distribution agreements, was green-lit by the CMA as it did not amount to ‘material’ influence – a lower threshold.  These clearances have adapted to the strategic barriers that such agreements may create, but have not escaped criticism for under-enforcement. Hyperscalers have preferential access to these AI models, to the detriment of smaller competitors in the downstream market, and can use their ‘relationship’ to cement users in their wider product/service ‘ecosystems’ (tying and bundling), thereby allowing for higher pricing. Regulators have responded by revisiting their thresholds and expanding their investigations into ecosystem practices to reduce such possibilities. Some of these concerns, along with the ecosystem structure, are briefly encapsulated in Figure 1. 

Figure 1: The FM/AI Ecosystem and the Vertical Stack

Layer 5 - Applications

ChatGPT, Claude.ai, Copilot,  Gemini, Perplexity, Enterprise SaaS, AI agents

Competition concern: self-preferencing via vertical integration, bundling with OS, browser and office suites

Layer 4 – Foundation Models

OpenAI, Anthropic, Google DeepMind, Meta (Llama), Mistral, xAI

Competition concern: API dependency with closed FM models (the first three), switching costs, hyperscaler investment, control over FMs

Layer 3 – Cloud and compute infrastructure

AWS, Microsoft Azure, Google Cloud, Oracle Cloud, CoreWeave

Competition concern: compute gatekeeping, exclusive partnerships with FMs built in with ‘agreements’ funding infrastructure, pricing discrimination

Layer 2 – Chip design

Nvidia (dominant), with fringe players AMD, Intel,  Google TPU, AWS Trainium, Meta MTIA. AMD competes directly with Nvidia in the Data Center CPU/GPU/DPU’s, with Intel a distant third.

Competition concern: Nvidia market power; hyperscaler vertical integration into custom silicon

LAYER 1 - SEMICONDUCTOR FABRICATION

TSMC (dominant), Samsung, Intel Foundry

This article seeks to provide a quick perspective of the evolving competition law concerns surrounding AI FMs – more specifically the dependency (strategic agreements) of (between) FM firms on (and) Hyperscalers and the degree of competition in this limited ‘market for inference’. Defining the relevant market as such is bound to attract much challenge and effort for a competition agency, and some may argue for an ‘ecosystem’ definition as a better approach. An alternative would be to have an ex ante regulatory framework. India’s proposed Digital Competition Law Bill, however, continues to languish. Understanding the degree of competition between FMs and the downstream consequences requires a look at both structural and behavioural factors, which are examined below. This essentially is a part of a Section 4 analysis, as it focuses on the ability of FMs to act independently of competitive forces using both the benefits of a first mover as well as ecosystem advantages. Market power is sought to be gauged using structural and strategic barriers to entry and its reward in terms of higher market share and pricing power.

II.            Foundation Models and the Structural Barriers

Examining structural factors such as capex and trends in innovation and entry provides a good understanding of barriers and the nature of competition ‘within’ and ‘for’ the market. Of course, sunk costs cannot be ignored. Large capex requirements, expected to touch $ 900 billion this year, create large economies of scale, yet low marginal cost in end-stream deployment for FMs. They also enjoy large economies of scope, as it is more efficient to build and train one foundation model that can then be deployed for multiple AI tasks downstream rather than train a model on one specialised task. Recovery of such large capital costs will need sufficiently high market shares and prices, both of which will require adequate market power, tempting the use of anticompetitive mechanisms and/or a race typical of platforms ( low or zero pricing), which will only see some earn adequate returns on investment, forcing exit and causing entry barriers.  Strategic ties with Hyperscalers can make a critical difference to achieve this end, and such relationships need not just reflect dependence. These factors, along with feedback loops, create additional strategic barriers and reduce contestability and the natural tendency for high concentration. But drawing conclusions about first-mover advantage for early FMs will need to account for innovation and switching costs, which can have consequences for entry.

Before we proceed, a quick understanding of the categories of FM versions (based on OSAID) will be useful: 1. Closed AI – access only through API (weights, training data and source code not available) includes OpenAI's GPT-4o, Anthropic's Claude versions, Google's Gemini Ultra; 2. Open models (all three available but subject to restrictions) ; 3. Open-weights(only weights are fully downloadable, with restrictions on others) include Meta's Llama series, Mistral AI models, Google's Gemma; 4. Open-source AI (all available freely)  includes OLMO, Amber, etc.  These categories, and the open-closed binary used in the charts that follow, are necessary simplifications. Scoring eleven leading models across eighteen legal, economic and governance variables, Schrepel and Potts find that most cluster in the middle of the openness spectrum - Llama 3 (open weight) and GPT-4 differ by a mere two points out of thirty-six - and caution against treating openness as binary (Schrepel and Potts, 2025). If models labelled open are only partially open, the restraint they exert on closed-model pricing will be weaker than their number suggests and strategic complementary and vertical relationships/integration can play a significant role.

Analysing Entry and Innovation: Entry data allows us to understand the extent of barriers to entry and competition in the market. The larger the number of new firm entrants, the greater the degree of competition. Chart 1 presents a time series frequency chart of the progression of FMs (compute ≥10^23) quarterly since 2019 Q4. This data is sourced from Epoch AI (the “known and unconfirmed compute-intensive models” large-scale subset) and has a different categorisation than the above four, which are collapsed to two (Open weight includes OSAID categories 2-4), and Closed relates to OSAID category – 1. Epoch AI categorises models with training compute ≥10²³ FLOP as ‘large-scale’ models — closely related to FMs, and to the General-Purpose AI (GPAI) models of the, which are presumed to carry systemic risk at ≥10^25 FLOP.  Charts 1–4 report two definitions together: (A) models with confirmed, disclosed training compute ≥10²³ FLOP (318 models); and (B) additionally, models Epoch identifies as compute-intensive, but whose compute is undisclosed (526 models in total, from 138 developers). Closed-weight labs rarely disclose compute, so (A) can undercount closed models.

We can broadly divide the trend into three periods: 1. Early phase - Domination by a few big tech companies (2019 Q4 – 2022 Q1)  - namely, OpenAI, Google, Meta, DeepMind, presumably due to capex and compute barriers; 2. Expansion & Entrant Explosion phase (2022 Q2 – 2023 Q4) - New entrants aided by venture capital and an explosion of ‘open source’ models; 3. Maturation & Incumbent Model Velocity phase (2024–2026) – a normalisation in the growth rate dominated by the top few.  Charts 2 to 4 show who releases these models and under what access terms. This shows a significant number of new entrants and the release of new models, indicating dynamic competition. That the number of ‘open weight’ FMs has exploded should ideally act as a competitive restraint on closed models and hence pricing, but as noted above, they are not really so ‘open’. But we also see signs of the market settling down with new entrants decreasing sharply. Along with other structural barriers, feedback loops among the consolidated players can make it more difficult for a native Indian FM entrant.

Explanatory note: Bars (left axis) count firms reaching ≥10²³-FLOP scale for the first time each quarter — dark blue= entry with a disclosed-compute model (definition A), light blue = entry via a compute-intensive model whose FLOP is unconfirmed (definition B); the line (right axis) is the cumulative frequency of distinct firms, reaching 138 by 2026 Q3. Entry accelerates sharply from 2023, but a rising firm count overstates rivalry: reaching ≥10²³ once is not sustaining frontier output, and many entrants release a single model. Source: Epoch AI, Data on AI Models (large-scale subset), retrieved 20 Aug 2026.

Explanatory note: Quarterly large-scale releases split by whether the developer was a first-time entrant that quarter (orange) or an incumbent already at scale (blue). Incumbents account for 350 of 526 releases and their share grows over 2024–2026: the innovation frontier is being advanced mainly by established labs iterating, not by new entry. Counts use the inclusive population (B). Source: Epoch AI, as Chart 1.

Explanatory note: Releases are split by open vs closed weights and, within each, confirmed ≥10²³ (solid) vs unconfirmed compute-intensive (pale). Among the confirmed (A), open dominates — 224 open vs 93 closed; but closed frontier models sit overwhelmingly in the unconfirmed category (156 of 249 closed) because closed labs rarely disclose compute. It’s only in the inclusive view (B) the totals are nearly even — 254 open vs 249 closed. ‘Open dominance’ by release count is thus substantially a disclosure number, and says nothing about capability, usage or revenue. 23 models with undetermined accessibility are excluded. Source: Epoch AI, as Chart 1.

Explanatory note: The fifteen developers with the most large-scale models (inclusive definition B) are categorised by open-weight vs closed / API / unreleased. Open-weight output is led by Alibaba, Meta AI, NVIDIA and DeepSeek; closed output by OpenAI, Anthropic, Google DeepMind and Amazon — a very striking firm-specific split. Source: Epoch AI, as Chart 1.

Of concern are the deal records. Tracking of AI-related investments, alliances and acquisitions shows activity rising from just a couple in 2022 to 68 in 2025 and 61 in the first seven-and-a-half months of 2026 alone (Chart 5). Agreements related to infrastructure deals - compute capacity, data centres, and increasingly the physical layer beneath them- rose from about a quarter of tracked deals in 2025 to a clear majority in 2026 (Chart 6).  These deals now go beyond compute and data-centre capacity to dedicated power generation (including gas and small-modular-nuclear supply), liquid cooling, copper and electrical conductors, and optical interconnects. This is the same vertical stack set out in Figure 1, now being secured through contract and acquisition. For a prospective entrant - including an Indian FM developer - the barrier is therefore not only the headline capex but that the complementary inputs at each layer are being locked up by incumbents that already control the layers above, a barrier that is both structural and strategic.

Explanatory note: Deal counts by year using Mogin’s own category labels (‘Infrastructure / Materials’ merged into Infrastructure). Source: Mogin Law A.I. Deal Table (moginlawllp.com/mogin-law-ai-deal-table), retrieved 20 Aug 2026; Year 2019 only partially complete in the source.

Explanatory note: The Mogin Law A.I. Deal Table is a non-exhaustive list. Deal values are frequently undisclosed and some entries are reported or unconfirmed. It is used here for the direction and composition of deal-making, not precise totals. Because many values are undisclosed and a few mega-deals would dominate any monetary aggregate, Charts 5–6 use deal counts rather than summed values. Source: Mogin Law A.I. Deal Table, retrieved 20 Aug 2026.

III.          Behavioural Factors 

In behavioural factors, pricing and market share trends are considered. They both serve as a reasonable indicator of market power. While the entry data indicate a competitive market in terms of availability of demand-side substitutes, the number of players need not indicate adequate competitive restraints. To start our discussion, while FMs will need to generate revenue to recoup the large capex, current financials seem to paint a different picture. Revenues generated from AI-based services are approximated at $220 billion but, while growing fast, have proven insufficient to justify the capex expected to touch $900 billion this year. Currently, the ability of consumers to switch between priced and alternate-priced models and free FMs - as the entry of multiple players enhances demand-side substitutability, firm strategies aimed at poaching each other’s clients in a zero-sum game, and slower expansion in width and depth of AI usage all contribute to both revenue growth dampening and increasing within-market competitive restraints. But while this gap is not in itself evidence of monopoly rents today, this large spending far ahead of revenue may actually sharpen the concentration concern. The higher the capital hurdle relative to current revenue, the fewer the firms that can clear it without Hyperscaler funding, and the stronger the incentive to defend margins through pricing and contractual lock-in once scale is reached. This may also explain the rapid slowdown in new entry seen earlier. The following charts provide market share and pricing dynamics which can help provide a deeper perspective. 

Data is sourced from OpenRouter, which captures only the routed-API segment it represents. Market share trends (Chart 7) reveal that the market is clearly concentrated among a few firms; when revenue is used as the metric, even more so than suggested by the entry data.

Going by this data, and reading it only as the routed-API segment it represents, market share trends (Chart 7) reveal that the market is clearly concentrated amongst a few firms; when revenue is taken as the metric, even more tightly than suggested by the entry data.  

Source: OpenRouter, retrieved August 2026. OpenRouter captures developer/API-routing traffic only — it excludes first-party app and direct-enterprise usage (ChatGPT, Claude.ai, the Gemini app), so it over-represents open and switchable models and cannot be read as whole-market share. If anything, the bias runs against the finding: since the sample over-represents open and switchable models, revenue concentration in the wider market is likely to be higher, not lower.

Market share trends over time are a better reflection of competitiveness. Chart 8, sourced from OpenRouter, is three images merged to reflect the dates and colour codes read off a dynamic source chart, so there may be slight image adjustments, but the numbers match the source perfectly. To be noted – this is the market share of the same firms as LLM’s – in the area of text generation – and hence these figures must only be seen as a proxy  (the "LLM" layer of the FM stack). Also, as noted earlier, OpenRouter data captures adoption rather than model quality, covers only traffic routed through OpenRouter — not the whole market and not usage on a provider's own first-party API. Given the nature of an API router, it therefore over-represents open-weight and readily switchable models relative to closed frontier models whose volume flows largely through their own APIs

Chart 8 : Changing token market shares -  August 2025 to August 2026

The numbers indicate a shift from a Google-dominated market in 2025 toward a much more competitive and fragmented market in 2026, with DeepSeek making the largest gain. While this data is subject to OpenRouter’s traffic, the trends are telling - market shares are not stagnant, reflecting the earlier entry and innovation statistics.

But Chart 9 provides pricing distribution, which shows a big asymmetry between closed- and open-source FMs. Pricing is as much a function of costs, which have fallen steeply, as of market power.

Epoch AI finds that the cost of reaching a fixed performance milestone has fallen by roughly 9x to 900x per year depending on the task, with the sharpest declines most recent.

While this can explain price declines, it cannot account for the large asymmetry. Can qualitative factors provide an explanation – are we seeing a differentiated price -quality spread? On Epoch AI’s Capabilities Index (ECI), the most capable open-weight models have lagged frontier closed models by an average of only about four months (roughly 8 ECI points) since January 2026, and Chinese models have trailed the US frontier by an average of seven months since 2023.[3] So a quality lead measured in months cannot on its own account for a price difference of this order; the explanation has to be found elsewhere.

Part of it is ordinary price discrimination. Second-degree discrimination is visible in the tariff itself - output tokens are priced several times above input tokens, commonly four to eight times among the frontier closed models, reflecting the greater compute that generation consumes, while tiered versions of the same model invite users to self-select. Third-degree discrimination operates across segments: the leading firms, particularly Anthropic,  concentrating on enterprise, coding and agentic workloads where willingness to pay is highest; published prices among closed models of the same vintage differ by as much as twelve times, while measured quality differs by a few percentage points (Value Add VC, June 2026). Further, list prices are also not realised prices – several other usage factors can affect costs. Once these are allowed for a residual remains, and it tracks vertical integration. Closed US FMs are embedded in the service offerings of their Hyperscaler partners - Word, Excel and PowerPoint through Microsoft Copilot, and Workspace through Gemini - which delivers deeper workflow integration, enterprise security and support commitments, and therefore better client response and satisfaction. This is an advantage that Chinese open-weight models cannot presently offer, whatever their benchmark scores, and geopolitical alignment reinforces the pairing of US FMs with US Hyperscalers.[4] From a competition perspective, it means the price asymmetry in Chart 9 should be read as evidence of switching costs and vertical integration rather than of model quality alone - which strengthens the strategic barrier and dependence concern developed earlier. Two points to be noted – Epoch mentions that closed labs may keep their most capable models unreleased, so the true quality gap could be wider, and the ECI measures capability ceilings rather than reliability in deployment.

Explanatory note. Source - Layer3Labs pricing index[5]: Open-weight models cluster at $0.10–$0.30 per million tokens (Mistral Large 2 the outlier at $2/$6); i closed models at $1.25–$15.00. Anthropic’s Opus tier ($25 output) is excluded as it would compress the scale

IV.          Conclusion 

On a positive note, we see procompetitive trends in innovation rate and new-entry frequency, with lower pricing points and, on Epoch’s capabilities index, a closed-to-open quality gap now measured in months rather than generations. But while options remain, price pressure may persist among consumers with a lower willingness to pay; enterprise solutions will face cost challenges due to strategic tie-ups. Vertical integration and resulting self-preferencing are factors that the CCI can act against within the existing regulations, and reverse acquihires may be viewed through the ‘material control’ prism provided by The Competition Amendment Act (2024). When examining market power, while acknowledging that open-source can play a critical role in restraining pricing power, behavioural factors, switching costs, and brand trust issues may still confer significant market power on closed-end models, as reflected in price asymmetry. The ecosystem integration: the embedding of closed US FMs in the Hyperscalers’ own productivity suites, and the geopolitical alignment now formalised in initiatives such as Pax Silica and informally held together through threats^, together sustain a price premium that the narrowing quality gap alone would not support. Price discrimination, while a normal feature in differentiated markets, can still attract the attention of the CCI when considered along with these potential leveraging lock-ins. Diversification of cloud compute and hardware sources will be critical but is unlikely in the immediate future and will require systematic intervention by regulators to democratise data. But government efforts to finance indigenous FMs and to develop data, data centres and compute will be critical in lowering these entry barriers, even if they cannot dismantle all of them. India’s accession to Pax Silica may ease access to trusted compute and equipment, yet it does not by itself address the deeper dependence mapped in Charts 5 and 6 — the progressive locking-up of every layer of the stack, from data centres down to copper, by the same firms that dominate the layers above. That, rather than the number of models released, is the question CCI and other regulators in India will have to confront.

From an ex-ante perspective, the Digital Competition Bill, still unenacted, may be better placed to tackle these issues. Hyperscalers would be categorised as Systemically Significant Digital Enterprises, and the clause related to the Associate Digital Enterprise mechanism can pre-empt self-preferencing, tying and restrictive data use to a group entity such as an FM arm that benefits from data collected through that service. These obligations are based on status and hence do not touch merger review. Foundation models could be listed as Core Digital Services. It is encouraging that the Standing Committee on Finance is pressing for finalisation of the Bill.

Notes:

[1]  The term ‘horizontal’ partnerships is used not from a collusion perspective but rather to illustrate the relationship between data / infrastructure and algorithm developing AI firms which are all needed to deliver FMs. Wong discusses how such talent hires can still be termed as ‘asset’ acquisitions and be subject to Section 7 of the Clayton Act which oversees mergers and acquisitions. Such agreements now span critical infrastructure such as power and cooling. See Mogin Law ‘From Data Centre’s to Copper: AI Infrastructure Race Goes Deeper’, available at   https://moginlawllp.com/from-data-centers-to-copper-ai-infrastructure-race-goes-deeper/

[2] See Market study on Artificial Intelligence and Competition, CCI, Available at: https://www.cci.gov.in/economics-research/market-studies/details/47/0; UK CMA, ‘AI Foundation Models: update paper’ , 11 April 2024, Available at https://www.gov.uk/government/publications/ai-foundation-models-update-paper ; US FTC, FTC Issues Staff Report on AI Partnerships & Investments Study, January 17, 2025 https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-issues-staff-report-ai-partnerships-investments-study;  European Commission, Commission takes note of the withdrawal of referral requests by Member States concerning the acquisition of certain assets of Inflection by Microsoft , 18 September 2024

[3] Jack Edwards and Luke Emberson, ‘Open models lag state-of-the-art closed models by 4 months’, Epoch AI Data Insight, 2026, https://epoch.ai/data-insights/open-closed-eci-gap; Epoch AI Data Insight, 2026, https://epoch.ai/data-insights/us-vs-china-eci;  The Epoch Capabilities Index (ECI) is Epoch’s aggregate measure of model capability.

[4] Pax Silica is a US State Department initiative declared in December 2025 to coordinate ‘trusted’ supply chains spanning critical minerals, energy, advanced manufacturing, semiconductors, AI infrastructure and logistics. Founding signatories included the United States, Japan, South Korea, Singapore, the Netherlands, the United Kingdom, Israel, the United Arab Emirates and Australia; membership had expanded to around two dozen partners, including India and the European Union, by the June 2026 summit. See https://www.state.gov/pax-silica.

[5] Layer3Labs, ‘AI Model Pricing Compared: Cost Per Token’ (verified 28 July 2026); each row source-linked to the vendor’s official pricing page. https://www.layer3labs.io/ai-model-pricing