AI Companies: The Complete Guide to the Global AI Industry

Search for "AI company" and you are asking one of the broadest questions in modern technology. The label now covers silicon designers, hyperscale cloud providers, model labs, deployment platforms, agent builders, generative media studios and vertical specialists in law, healthcare, finance and robotics. This guide breaks the industry down layer by layer, names the businesses that matter in each, explains how they earn revenue and shows how to choose the right AI partner for your organisation — whether you are a UK enterprise buyer, an in-house digital team or an ambitious challenger brand looking to stand out.

iCentric Agency works with clients across financial services, professional services, SaaS and consumer brands to design, launch and optimise AI-powered digital experiences. The observations below draw on that ongoing engagement, on the published disclosures of the major AI companies themselves and on the wider industry taxonomy that has emerged as the space has matured.

What is an AI company?

At its narrowest, an AI company is one whose core product is an artificial intelligence system — a model, a training or serving platform, or an application whose usefulness collapses without a learned model at its centre. At its broadest, the label is applied to almost any software business that has embedded machine learning into a feature set. The difference matters, because the two ends of that spectrum have very different unit economics, defensibility profiles and risk exposures.

A pure-play AI company typically has one or more of the following characteristics. It employs a disproportionate share of research and machine learning engineering talent relative to conventional product and platform staff. It spends heavily on compute — either buying GPU capacity from the hyperscalers or, at the largest scale, building its own data centres and buying accelerators directly. It ships product releases that are gated by model quality rather than by user interface work, meaning that big jumps in capability arrive on the cadence of research breakthroughs rather than sprint cycles. And it competes for benchmark performance, publishing evaluation results as a marketing surface in a way that traditional SaaS companies do not.

An AI-enabled company, by contrast, typically buys the underlying intelligence from someone else. It uses OpenAI, Anthropic, Google or an open-weights alternative through an API, layers retrieval and orchestration on top and applies its own product, distribution and domain expertise to package the result. That does not make it any less real as a business — much of the value in AI accrues at the application layer where the data, workflows and customer relationships live — but it does change the strategic posture. The AI-enabled company competes on product-market fit, workflow understanding and go-to-market execution; the pure-play AI company competes on the underlying technology.

There is an obvious commercial reason so many businesses now claim the AI label. Public and private investors reward it, buyers ask about it explicitly in RFPs, and talent markets pay a premium for teams involved with the category. The result is that the term is used loosely. For procurement and strategy purposes, we recommend asking two straightforward questions of any vendor calling itself an AI company. First, if their access to third-party foundation models disappeared overnight, would their product still function? Second, what proportion of their headcount is in ML research or ML engineering, and what proportion of their operating cost sits with a compute provider? Those two answers place a vendor accurately on the spectrum between pure play and AI-enabled, and they are far more informative than any marketing collateral.

The layers of the AI stack

The global AI industry has organised itself into a stack, and every AI company sits at one or more of its layers. Understanding the layers is the fastest way to make sense of the field and to identify where a given supplier competes.

At the base sits silicon and hardware. This is the world of NVIDIA, AMD, Intel, Broadcom, Google's TPUs, Amazon's Trainium and Inferentia, and a small group of specialist accelerator designers such as SambaNova, Cerebras, Groq and Tenstorrent. Companies at this layer earn revenue by shipping physical chips and the software that programmes them. The barriers to entry are enormous — fabrication access, driver ecosystems and developer familiarity — which is why the base of the stack remains the most concentrated part of the market.

One layer up sits cloud, GPU orchestration and inference infrastructure. This is where hyperscalers (AWS, Azure, Google Cloud, Oracle Cloud) meet a growing group of specialists. Neocloud operators such as CoreWeave, Lambda and Crusoe rent GPU capacity at scale to AI-native workloads. Platforms such as Together AI, Fireworks AI, Baseten, Modal, Replicate and Northflank sit above the raw GPU layer and turn training and inference into a developer-friendly API. Their businesses are built on utilisation, latency and developer experience.

Above the infrastructure comes the foundation model layer. These are the labs and companies that train the general-purpose large language models, vision-language models, speech models and image and video models on which the rest of the industry depends. OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral, Cohere, DeepSeek, xAI, Alibaba's Qwen team and a growing list of others all compete here. So do specialist model developers focused on a particular modality — for example ElevenLabs in voice, Midjourney and Black Forest Labs in images, Suno in music.

The orchestration, retrieval and agent framework layer is a younger stratum that has grown rapidly. It contains companies and open-source projects that make foundation models useful in production. Vector databases (Pinecone, Weaviate, Qdrant, Chroma), retrieval frameworks (LlamaIndex, LangChain), agent frameworks and evaluation platforms (Braintrust, LangSmith, Humanloop, Arize) all sit here. So do specialist orchestration platforms designed to let agents call tools, browse the web and execute workflows.

Above that sits the application layer: copilots, agents and vertical software that end users actually interact with. This is where most of the practical business value of AI is being captured today. Enterprise search platforms like Glean, legal automation companies like Harvey and Legora, customer service agent companies like Sierra and Decagon, finance and healthcare specialists — all live at the application layer. So do the generative media tools most people encounter first, from ChatGPT itself to Perplexity, Cursor, Notion AI and Synthesia.

Finally, wrapping around all of this, sit two supporting layers: data labelling, curation and evaluation (Scale AI, Surge AI, Mercor, Snorkel) and safety, security and governance (Cyera, Protect AI, HiddenLayer, Robust Intelligence). These layers are less visible but are increasingly important as AI systems move into regulated industries.

When we say "AI company" without qualification, we usually mean either the foundation model layer or the application layer, because those are the parts of the stack most people interact with. The full picture is broader, and if you are choosing partners, the layer they operate at tells you almost everything about their business model, their defensibility and their commercial risk profile.

Foundation model developers

The foundation model layer is where the story of the current AI era began and where the largest concentrations of research talent and compute continue to sit. Understanding the leading model developers is essential context for anyone evaluating suppliers further up the stack, because their choices ripple through every downstream product.

OpenAI remains the reference point for the industry. Founded in San Francisco in 2015 and originally structured as a research non-profit, it now operates as a capped-profit company with Microsoft as its principal strategic partner. Its GPT family of models powers ChatGPT — the fastest consumer product to reach hundreds of millions of users — as well as a wide range of enterprise and developer offerings. OpenAI competes on three fronts simultaneously: consumer subscriptions, developer APIs and enterprise deals negotiated at the executive level. Its scale gives it distribution advantages that few competitors can match, and its model releases set the pace for the rest of the market.

Anthropic, founded by former OpenAI researchers, positions itself as the safety-focused frontier lab. Its Claude family of models has become the default second option for enterprise buyers who want a credible alternative to GPT and has become particularly strong for coding, agentic workflows and long-context reasoning. Anthropic has secured strategic backing from Amazon and Google, giving it access to multiple hyperscale distribution channels. Its research publications on constitutional AI, mechanistic interpretability and safe deployment are influential across the industry.

Google DeepMind is the merged research organisation that combines Google Brain and the London-based DeepMind. Its Gemini family of models is deeply integrated into Google Search, Workspace, Android and Google Cloud, giving Alphabet a distribution moat that pure-play labs struggle to match. DeepMind's research pedigree — AlphaGo, AlphaFold, the reinforcement learning breakthroughs of the last decade — gives it a distinctive scientific culture, and its work through Isomorphic Labs on drug discovery is one of the most interesting applied research programmes anywhere in the industry.

Meta AI takes a different approach: it releases its Llama family of models with open weights, meaning the model parameters are available for download and self-hosting. This has made Llama the default choice for organisations that need on-premises deployment for regulatory or commercial reasons and for the wider open-source community that builds fine-tuned variants. Meta funds its model programme from its advertising business rather than from direct model revenue, which gives it a fundamentally different economic model to the pure-play labs.

Mistral AI, founded in Paris, is Europe's most prominent frontier lab. It ships a portfolio that spans open-weight and proprietary models, and its focus on efficiency has produced small models that punch above their weight on cost and latency. Mistral's local roots have proved to be a commercial asset with European governments and enterprises that want a home-continent supplier, and its distribution partnerships now span Microsoft Azure, Cisco and multiple sovereign cloud programmes.

Cohere, headquartered in Toronto, focuses squarely on enterprise. It builds models optimised for retrieval-augmented generation and secure deployment in customer environments, and its go-to-market emphasis on regulated industries has made it a natural fit for banks, insurers and government agencies. Its Command and Embed model lines are widely used inside customer-built assistants.

DeepSeek is the best known of the Chinese open-source model developers. Its releases have consistently demonstrated frontier-class performance at a fraction of the compute assumed to be necessary by Western labs, and its open-weight approach has made it hugely influential inside the research community. Alongside DeepSeek sit other Chinese labs — Alibaba's Qwen team, Baichuan, Moonshot AI (Kimi), MiniMax, Zhipu (Z.ai) and Tencent's Hunyuan — each of which is shipping increasingly competitive models.

A new generation of research-heavy labs is also emerging. Safe Superintelligence Inc., founded by Ilya Sutskever, is pursuing a single research programme aimed at safe frontier systems. Thinking Machines Lab, founded by Mira Murati with a large cohort of former OpenAI researchers, is another closely watched contender. Reflection, headquartered in New York, is building open-source frontier models and has been positioned as an American counterweight to DeepSeek. xAI, Elon Musk's lab, is developing the Grok family of models and is aggressively building out its own supercomputing capacity.

When enterprise buyers ask us which foundation model provider they should build on, the honest answer is that most large-scale programmes end up multi-model. The right approach is to keep model choice as an abstraction inside your orchestration layer, evaluate models against your own workloads on a regular cadence, and treat any single provider as replaceable at short notice.

AI infrastructure and compute providers

Behind every model release sits a very large compute bill and a set of infrastructure companies that make training and inference possible. This layer of the AI industry is where an increasing share of the total spend now lands, and it is where several of the industry's most important businesses have been built.

NVIDIA is the fulcrum. Its H100 and B200 accelerators, its NVLink interconnects and, above all, its CUDA software ecosystem have made it the default choice for training and serving large models. The company's market position derives less from any single chip and more from a decade-long investment in software libraries and developer relationships that competitors are still trying to replicate. Every foundation model lab of consequence is an NVIDIA customer, and every hyperscaler builds its GPU capacity around NVIDIA hardware alongside its own accelerators.

AMD has emerged as a credible second source at the chip layer, with its Instinct series accelerators gaining traction inside hyperscale customers. Intel continues to compete with Gaudi accelerators. Alongside them sits a group of specialist silicon designers targeting inference rather than training: Groq, whose LPUs deliver very high token throughput; Cerebras, whose wafer-scale processors offer distinctive economics for certain workloads; SambaNova, which sells fully integrated systems and models to enterprise and government customers; and Graphcore, the UK-founded IPU designer now operating under Softbank ownership.

One layer up sits the AI data centre and neocloud category. Crusoe, based in Denver, has built a business around locating high-density GPU data centres near cheap and often stranded energy sources. CoreWeave and Lambda are the largest independent GPU cloud providers, renting NVIDIA capacity to AI customers who cannot get it in the volumes they need from the hyperscalers. Nebius, Applied Digital and a growing list of regional operators serve similar demand across other markets.

Above the raw compute layer sits a rapidly consolidating group of deployment and serving platforms. Databricks, headquartered in San Francisco, has evolved from a data lakehouse company into a full data and AI platform, offering everything from data storage and analytics through model training, fine-tuning and deployment. Its acquisition of MosaicML gave it a strong training story, and its Mosaic Agent Framework is now used across Fortune 500 customers. Snowflake, its closest competitor, is executing a similar strategy from a different starting point.

Together AI and Fireworks AI operate very fast inference platforms tuned specifically for open-weight models — Llama, Mistral, Qwen, DeepSeek and others — and have become preferred serving layers for companies who want to avoid the costs and rate limits of the frontier labs' first-party APIs. Baseten offers a similar service with a strong developer experience story around cold-start latency and custom model deployment. Modal, Replicate and RunPod each occupy adjacent positions in the same market.

Northflank, headquartered in London, has built a deployment platform designed for AI workloads that spans training, inference and long-running agent processes. Its sandboxing capabilities make it useful for running untrusted AI-generated code, which has become a real problem for companies building autonomous coding agents. Northflank is one of a group of UK-headquartered infrastructure companies whose customer base is now heavily international.

Sitting alongside the deployment platforms are agent runtime and orchestration companies. LangChain and LlamaIndex started as open-source frameworks and have grown into commercial platforms that provide evaluation, monitoring and deployment for LLM applications. Braintrust, Humanloop and Arize AI compete in the LLM observability and evaluation category. Pinecone, Weaviate, Qdrant and Chroma compete in vector search.

When we advise clients on infrastructure choices, the pattern is usually the same. Frontier applications hit hyperscaler APIs first because time-to-first-value is short. As volumes grow, they migrate cost-sensitive workloads to specialist inference platforms. And regulated customers who need private deployments increasingly pull models fully in-house, hosted on Kubernetes clusters running on top of a neocloud or their own data centre capacity.

Enterprise AI application companies

Most of the practical business impact of AI shows up at the application layer, where models are packaged into workflows that end users can actually operate. This is also where the most interesting new companies of the current cycle are being built.

Enterprise search and knowledge agents are led by Glean, headquartered in Palo Alto. Glean connects to a company's SaaS estate — Slack, Google Drive, SharePoint, Jira, Confluence, Salesforce, Zendesk and dozens of other sources — and turns the collective corpus into an assistant that employees can ask natural-language questions and delegate tasks to. Its success has spawned a growing category of competitors and platform features from the SaaS incumbents.

Legal automation has become one of the most defensible verticals for AI, because the workflows are structured, the outputs are text-heavy and the customers are willing to pay for accuracy. Harvey, in San Francisco, has become the standard for large law firm deployments in the United States and, increasingly, the United Kingdom. Legora, headquartered in Stockholm, is Harvey's most credible European competitor, with strong penetration into UK, Nordic and continental firms. Both companies are moving rapidly from document review and drafting into transactional workflow automation.

Financial services have thrown up a distinct set of AI application companies. Rogo builds an assistant used by investment bankers and public-market investors to interrogate filings, transcripts and internal research. Hebbia offers a similar workflow for asset managers. Both compete against internal builds from the major banks, which are themselves training in-house models.

Healthcare application companies include Abridge, whose ambient scribe listens to patient conversations and produces structured clinical notes; OpenEvidence, an AI-powered medical search engine for clinicians; and Nuance, now Microsoft-owned, which occupies the same category with far larger installed base. EliseAI, originally focused on lettings, has expanded into healthcare workflows.

Customer service is being reshaped by Sierra, founded by Bret Taylor, whose agents handle inbound conversations across chat, voice and email; and by Decagon, whose deployments span retail, fintech and travel. Ada and Cresta occupy adjacent positions. The general pattern is that the best customer service AI companies replace tier-one contact centre work outright, escalate to humans on a much smaller share of contacts and integrate deeply into CRMs and knowledge bases.

Sales and go-to-market AI companies include Clay, whose enrichment and outbound platform has become the default tool for growth teams, and Listen Labs, which automates customer research interviews. Gamma — which is technically a productivity tool — has become an unlikely GTM darling because it lets sellers assemble narrative decks quickly.

Data labelling and evaluation is the quieter foundation on which many of these application companies depend. Scale AI remains the largest player, but Surge AI, Mercor and Snorkel all occupy significant positions, and the labelling category has evolved from crowd-sourced click work into specialist expert networks that fine-tune models for specific domains.

When an organisation asks us where to start with enterprise AI, we usually recommend one of two entry points. Either deploy an enterprise search assistant across the whole workforce and let horizontal productivity gains accrue while you learn — Glean is the archetype here — or select one high-value vertical workflow, deploy a specialist application (Harvey, Sierra, Abridge, Rogo depending on your industry) and use it as a wedge for broader change. Both paths work; both require serious change management rather than a purely technical project.

Generative media companies

Generative media is where AI has crossed most decisively into culture. It is also a category where a small number of well-funded companies are competing on model quality, product design and creator relationships.

Voice AI is dominated by ElevenLabs, headquartered in New York with a significant London engineering base. Its text-to-speech and voice cloning models are the reference points for the category, and its recent push into voice agents has extended the platform into real-time conversational applications. PlayHT, Resemble AI and Cartesia compete in the same market with different technical bets.

Music generation has emerged as its own category, led by Suno, based in Cambridge, Massachusetts, and Udio. Both companies have generated cultural attention — and predictable legal attention from rights holders — by making high-quality full-track music generation available to non-musicians. Riffusion occupies an adjacent position.

Image generation is a fragmenting market. Midjourney, still independent and famously bootstrapped, continues to command deep creator loyalty. Black Forest Labs, founded by the team behind Stable Diffusion, ships the Flux family of models and has quickly become the preferred provider for commercial image work. Ideogram has carved out a niche on typography. Krea blends its own models with third-party model access into a creative product. Fal provides generative media infrastructure that many of these companies build on top of. Stability AI, headquartered in London, is under new ownership after a turbulent period and continues to ship open-weight models.

Video generation has moved rapidly from novelty to near-production quality. Runway, in New York, is the industry's most credible video generation company for creative professionals. Luma Labs ships the Dream Machine family. Google's Veo, OpenAI's Sora and MiniMax's Hailuo all compete on model capability. Chinese entrants Kling and Seedance have advanced quickly and are widely used inside creator communities.

Corporate video and avatar is a separate market. Synthesia, headquartered in London, has become the dominant provider of AI avatar video for enterprise learning, internal communications and localisation. HeyGen, based in Los Angeles, occupies a broadly similar position with a more marketing-led product. Both companies have moved rapidly from single-shot avatar generation into full studio workflows.

Design and productivity has produced perhaps the most surprising winners of the cycle. Gamma turns prompts into full presentations, decks and websites. Notion has become one of the more successful cases of an incumbent SaaS company adding AI as a first-class feature. Adobe's Firefly integration into Creative Cloud represents the incumbents' answer to the entire wave.

For marketing teams, the generative media category is where the direct opportunity is most obvious — content production costs collapse, personalisation surfaces expand and net-new formats become viable. It is also where the intellectual property, brand safety and disclosure questions are most acute. We advise clients to introduce generative media tools with clear guardrails from day one: approved model providers, prompt libraries, review workflows and disclosure standards.

AI coding and development companies

A distinct category of AI companies has emerged around software development, and the pace of change here has been faster than anywhere else.

Cursor, built by Anysphere, has become the dominant AI-native code editor, displacing significant share from traditional IDEs in a very short period. Its value proposition combines rapid inline completion with agentic workflows that can plan and execute multi-file changes. Windsurf (originally Codeium) competes at a similar layer.

Cognition is the company behind Devin, an autonomous coding agent that can pick up tickets, run tests and open pull requests. It represents a different bet from Cursor: rather than augmenting the developer, Devin aims to replace a share of routine engineering work outright. Poolside AI, headquartered between Paris and New York, is another well-funded frontier lab focused specifically on code models.

Replit has evolved from a browser-based coding environment into an AI-first application builder, and Lovable, headquartered in Stockholm, competes head-on in the AI app and website builder category. Both are aiming squarely at the non-developer building simple software.

Microsoft's GitHub Copilot remains the largest single AI coding product by seat count, and Google's Gemini Code Assist is deployed inside a very large enterprise base. The incumbents' distribution advantages are considerable, but the pace at which the specialist AI coding companies have shipped has kept the market genuinely competitive.

For engineering leaders, the practical question is no longer whether to introduce AI coding tools but which tool to standardise on, how to evaluate its impact on quality and how to handle the security implications of AI-generated code entering the codebase. Every serious rollout we have supported has included an evaluation harness for the tool's output, updates to code review policy and clear guidance on where autonomous agents are and are not permitted to act.

Vertical AI companies

Beyond the general enterprise categories, a large group of AI companies is targeting specific industries and building deep, domain-specific products.

In healthcare and life sciences, alongside the applications already mentioned, sit companies working on drug discovery and molecular design. Isomorphic Labs, Alphabet's spin-out from DeepMind, is using descendants of AlphaFold for pharmaceutical research. Chai Discovery operates in the same space. BenevolentAI, headquartered in London, has been a long-standing player, and Recursion combines automated laboratory work with machine learning for target discovery.

In legal, beyond Harvey and Legora, sit specialist companies including Ironclad, Robin AI and Luminance — the last two both UK-founded. Each attacks a slightly different slice of the legal workflow.

In financial services, alongside Rogo and Hebbia sit AlphaSense in market intelligence, Uplimit in workforce enablement and a growing group of quantitative trading firms with significant in-house AI capabilities. Cyera operates in AI-driven data security, an increasingly important category as companies deploy generative AI over sensitive corporate data.

In robotics and physical AI, a new generation of companies is applying foundation model techniques to embodied systems. Physical Intelligence, in San Francisco, is training generalist robotics models. Skild AI, in Pittsburgh, is building AI systems for humanoid and industrial robotics. Figure AI and 1X are developing humanoid platforms. Wayve, headquartered in London, is applying end-to-end learning to autonomous driving and has partnerships with several major automakers. World Labs, co-founded by Fei-Fei Li, is developing spatial AI models that reason about three-dimensional environments.

In defence and national security, Anduril and Palantir are the most prominent examples of AI-heavy companies serving government customers, alongside a growing group of specialist entrants. This category is politically sensitive but commercially significant, and it is where a good deal of the most demanding real-time AI work is now being done.

Each vertical AI company sits somewhere on a spectrum between building its own models and leveraging third-party foundation models with domain-specific data and workflows. The best of them combine both: they hold proprietary data assets that give them an edge, they fine-tune or train their own smaller models on top, and they use frontier models where the extra capability is worth the cost.

The UK AI company landscape

The United Kingdom has an AI industry that punches significantly above its weight relative to the size of its economy, and understanding it matters for any UK-based buyer or investor.

Google DeepMind is the anchor of the London AI research cluster and remains one of the most influential research organisations anywhere in the world. Its work on protein structure (AlphaFold), reinforcement learning, materials science and now Gemini has shaped the entire industry, and its presence in King's Cross has made London a magnet for research talent.

Isomorphic Labs, spun out of DeepMind, is applying that research to pharmaceutical drug discovery in partnership with major pharma companies. It represents one of the clearest examples of an applied AI company built on top of a research organisation's intellectual capital.

Synthesia, headquartered in London, has become the dominant enterprise AI video platform, with tens of thousands of business customers using its avatars for learning content, internal communications and localisation. It is one of a handful of UK-founded AI companies to have achieved genuine category leadership at global scale.

ElevenLabs, while now headquartered in New York, has significant London engineering and business presence and is often described as an Anglo-American company. Its founders' background and much of its research team sit in the UK.

Stability AI, headquartered in London, retains a significant role in the open-weights image generation ecosystem under new ownership after a period of turbulence.

Wayve, also London-headquartered, is one of the most technically ambitious companies in autonomous driving, taking an end-to-end learning approach that is architecturally different from the sensor-fusion pipelines favoured by many US competitors.

Faculty AI is one of the UK's oldest applied-AI consultancies, with significant public sector and enterprise practice, particularly around decision intelligence.

BenevolentAI applies knowledge graph and machine learning approaches to drug discovery from its London base.

Northflank, mentioned earlier, is a London-founded infrastructure company whose deployment platform has become popular with AI-native customers.

PolyAI builds voice assistants used in enterprise contact centres from its London office.

Graphcore is the UK's most significant AI silicon company, designing intelligence processing units and now operating under Softbank ownership.

Alongside these headline names sits a much broader ecosystem of AI-first consultancies, specialist tooling providers and vertical application companies. Cambridge, Oxford, London, Edinburgh and Bristol all have distinctive strengths, supported by strong academic AI programmes and by the government-backed AI Safety Institute.

For UK organisations, the calculation of whether to buy from a UK AI company or a US alternative usually comes down to three factors: data residency and sovereignty; the availability of an on-the-ground team for enterprise support; and, increasingly, procurement policies that favour domestic suppliers for certain categories of work. On all three axes, UK-headquartered AI companies have got a much stronger story to tell than they did five years ago.

How AI companies make money

The business models used by AI companies are more varied than they first appear, and understanding them is essential when negotiating contracts or planning a build.

Consumption pricing, based on tokens, requests, minutes of audio or seconds of generated video, dominates the foundation model layer and much of the infrastructure layer. The advantage for buyers is direct alignment between usage and cost. The disadvantage is unpredictability: a viral product launch, a runaway agent or a poorly implemented retrieval loop can push costs sharply in the wrong direction, and finance teams struggle to plan against consumption bills. Effective governance requires per-tenant metering, budget alerts and abort logic inside applications.

Seat-based SaaS pricing dominates the application layer, particularly for productivity and knowledge tools where usage-based pricing would create adoption friction. Glean, Harvey, Legora, Cursor and many enterprise application companies sell primarily by seat, sometimes with usage caps and overage terms.

Enterprise contracts — annual or multi-year, with committed minimums and negotiated volume discounts — are how most large-scale AI spend is transacted. Enterprise buyers should expect meaningful discounts for multi-year commitments and should push for clauses that protect them from unilateral price rises when foundation model providers, in particular, adjust their upstream pricing.

Marketplaces and revenue share are how some platforms monetise — model marketplaces such as Hugging Face, agent marketplaces built on top of ChatGPT and voice agent marketplaces from ElevenLabs all use variants of this pattern.

Open weights with commercial services is the model favoured by Mistral, Meta, DeepSeek and others: the base models are freely available, and the company monetises through hosted APIs, enterprise support contracts, fine-tuning services or premium models that are not released openly. This is close to the classic open-source business model, adapted for an era where the cost of running the model matters at least as much as the cost of writing it.

Professional services and implementation provide meaningful revenue for many enterprise AI companies, particularly in regulated industries. Data labelling companies like Scale, Surge and Mercor are essentially services businesses with strong software components. Even pure product companies typically wrap their offerings in significant implementation support at the top end of the market.

When you evaluate a supplier, ask specifically how each of these revenue streams contributes to their business and how that shapes the incentives they will bring to the relationship. A company built on consumption pricing has a natural interest in encouraging expanded usage; a seat-based SaaS company has an interest in wall-to-wall deployment; a services-heavy company has an interest in extended engagement. None of these are wrong, but they change how the commercial conversation should be framed.

How to choose an AI company to work with

Selecting an AI supplier is a different exercise from selecting a conventional SaaS vendor. The technology moves faster, the risk surface is different, and the total cost of change is often higher than the sticker price of the software. Over dozens of client engagements, we have converged on a small set of evaluation criteria that reliably separate credible partners from marketing-driven imitators.

Model provenance and data handling should be the first question on the list. Which models does the supplier use? Are those models first-party, or are they served via a third-party API? Where is inference performed geographically? What happens to prompts, retrieved context and generated outputs — are they logged, cached, used for training, or destroyed on completion? Is there a documented data processing agreement, and does it align with the UK GDPR, the EU AI Act and any sector-specific rules that apply to your organisation?

Latency, uptime and throughput matter far more for AI systems than for conventional software, because slow responses collapse the user experience and hard rate limits from upstream providers can throttle your product. Ask for documented percentile latencies at your expected concurrency, not just average response times. Ask what fallback logic exists if a primary model provider degrades, and confirm that model choice is abstracted inside the supplier's platform.

Integration surface is the practical measure of how much work it will take to go live. Does the supplier support the identity provider, data warehouse, CRM, ticketing system and event bus you actually use? Are the integrations native or brokered through a third-party iPaaS? How mature are the webhooks, the audit logs and the administration API?

Developer experience is a strong proxy for engineering culture. Read the supplier's documentation as if you were about to build against it. Try their sandbox. Check the quality of their client SDKs. In our experience, the AI companies with the best developer experience tend to be the ones with the deepest technical benches and the most stable roadmaps.

Roadmap credibility and financial durability are harder to judge but essential for any multi-year commitment. Is the supplier profitable, or clearly on a path to profitability at their current cost base? How long is their runway? What is their retention profile — do customers renew and expand, or churn after a proof of value? Frontier AI companies are burning through capital at extraordinary rates, and buyers should stress-test any critical supplier against a scenario in which their next funding round is delayed.

Total cost of change vs. total cost of ownership is the calculation that anchors a serious procurement decision. The cost of adopting an AI supplier includes not just their subscription or usage bill but the integration effort, the change management, the data preparation, the training, the governance and evaluation infrastructure, and the risk of having to migrate to another supplier if the first choice does not work out. Buyers who focus only on the sticker price consistently underestimate the true cost of the programme.

When we support clients through supplier selection, we typically run a structured evaluation against three to five shortlisted vendors, using a scoring rubric that includes weighted questions across each of the criteria above and a live proof-of-value exercise against the client's own data. That process usually takes six to twelve weeks for a serious enterprise deployment and pays for itself many times over in reduced switching cost later.

Risks and considerations when relying on AI companies

Every significant AI programme brings risks that do not exist in conventional software procurement. Naming them upfront is essential.

Hallucinations and factual accuracy remain a defining risk of large language model applications. Even well-designed retrieval-augmented systems can generate confident-sounding text that is wrong in ways that are hard for non-expert users to spot. Mitigation requires careful prompt engineering, explicit grounding on trusted sources, evaluation frameworks that measure factuality on realistic tasks and human-in-the-loop workflows for high-stakes decisions.

Data leakage and IP exposure is a recurring concern with any external AI supplier. Employees pasting sensitive information into consumer AI tools has been the most common breach vector, and the solution is a combination of enterprise licences with proper data handling terms, network-level controls and staff training. When you buy from an AI company, insist on written commitments that your prompts and outputs will not be used for model training, and understand where retention windows sit.

Vendor lock-in and portability are harder in AI than in most other SaaS categories, because model behaviour is hard to reproduce across providers. A prompt tuned for one model may perform badly on another. Fine-tuned weights are typically not portable between providers. Mitigation requires designing your application layer so that model choice can be swapped, and maintaining an evaluation harness that lets you re-qualify a new provider quickly.

Regulatory exposure is expanding fast. The EU AI Act introduces risk-tiered obligations that apply extraterritorially to many UK organisations. UK sectoral regulators — the FCA, ICO, MHRA, Ofcom — are publishing increasingly specific guidance. Buyers should ensure their suppliers understand and can support the regulatory environment they operate in, and that contractual terms include appropriate warranties on compliance.

Bias, safety and reputational risk attach to any customer-facing AI system. A poorly grounded chatbot, a discriminatory decisioning model or a generative media system that produces embarrassing content can cause significant brand damage very quickly. Mitigation requires structured red-teaming before launch, ongoing monitoring in production and clear escalation paths when problems are detected.

Concentration risk is worth flagging separately. A very large share of the world's AI workloads flow through a small number of foundation model providers and an even smaller number of GPU manufacturers. Programmes that depend on any single component are exposed to outages, rate limits, price changes or geopolitical events beyond your control. Multi-model, multi-region and multi-provider designs are the sensible default for anything mission-critical.

Talent and skills risk rounds out the list. AI companies are competing hard for a limited pool of researchers and engineers, and the internal team you build to work with your suppliers will be a valuable target. Retaining that team requires meaningful investment in tooling, ongoing learning and interesting problems to work on.

None of these risks is a reason not to adopt AI. They are reasons to adopt it thoughtfully, with governance that matches the ambition of the programme.

How iCentric Agency helps organisations work with AI companies

iCentric Agency is a digital consultancy that helps ambitious organisations design, launch and optimise AI-native customer experiences. We are not a foundation model lab and not a systems integrator. We sit at the point where technology, brand and commercial strategy meet, and we work with clients who want to use AI companies effectively rather than build the underlying technology themselves.

Our work with clients across financial services, professional services, SaaS, retail and consumer brands typically covers four areas. Use case mapping — identifying where AI can meaningfully change a customer journey, an internal workflow or a product experience, and prioritising the opportunities that combine high value with achievable delivery. Vendor selection and technical due diligence — running structured evaluations against the criteria described above, from foundation model providers to specialist application vendors. Experience and content design — creating the on-page journeys, conversational flows, prompt libraries and content operations that make AI-driven products actually work for real users. SEO and generative engine optimisation — ensuring that as search behaviour shifts towards AI-mediated answers, our clients remain discoverable and cited by the AI companies whose products increasingly sit between brands and their customers.

We combine the disciplines of a brand and content agency with the technical depth of an AI consultancy, and we build small, senior teams around each client engagement. If you are evaluating how your organisation should work with AI companies — as a buyer, a partner or a competitor — we would welcome a conversation.

Frequently asked questions about AI companies

Which company is the biggest in AI? By almost any measure — model capability, revenue, distribution and cultural presence — OpenAI is currently the largest pure-play AI company. When you widen the definition to include the hyperscalers and hardware suppliers, NVIDIA and Microsoft become the most economically important AI companies in the world, followed by Alphabet, Meta and Amazon. Anthropic is the closest challenger to OpenAI at the foundation model layer.

What is the difference between an AI company and an AI-enabled company? A pure-play AI company develops the core intelligence itself — the models, the training infrastructure or the inference platforms — and its business collapses if that layer disappears. An AI-enabled company applies AI to a domain-specific problem using models supplied by others. Both are valuable, but they have very different economics, defensibility and risk profiles.

Are there any UK-based AI companies worth watching? Yes, and more than most buyers realise. Google DeepMind and Isomorphic Labs anchor the London research cluster. Synthesia and ElevenLabs are global category leaders in generative media. Wayve is one of the most technically interesting autonomous driving companies anywhere. Northflank, Graphcore, PolyAI, Faculty, Stability AI and BenevolentAI all occupy meaningful positions in their respective layers of the stack.

How should a mid-market organisation start working with an AI company? Start narrow and start with change management, not technology. Pick one workflow with high volume and high pain, deploy a proven application in that workflow — enterprise search, customer service, sales enablement, contact centre — and treat the first six months as a learning exercise rather than a scaling exercise. The organisations that succeed with AI treat it as an operating model change, not a software purchase.

How do open-source AI companies make money? Open-source and open-weights AI companies typically monetise through hosted APIs that make their models easy to run at scale, through enterprise support and fine-tuning services, through premium closed models offered alongside open ones, and through partnerships with cloud providers who resell their models. Mistral, Meta (indirectly, through advertising) and DeepSeek each demonstrate variants of this pattern.

Should we build our own model or buy from an AI company? For the vast majority of organisations, the answer is buy. Training a competitive foundation model is a research and infrastructure exercise measured in years and requires talent and compute that even large enterprises struggle to justify. Where organisations do build, it is usually smaller specialist models fine-tuned on proprietary data, running on top of open weights from Meta, Mistral or DeepSeek. That hybrid approach — buy the foundation, own the specialisation — is where most sophisticated programmes end up.

How do I make sure my organisation is not left behind? Focus on the fundamentals that compound over time: clean, well-governed data; a small internal team that is fluent in AI tooling; clear governance frameworks for procurement, evaluation and deployment; and a habit of experimenting continuously against a portfolio of use cases. Organisations that do those four things consistently tend to move faster than their competitors regardless of which specific AI companies dominate the next cycle.