Marcus Webb runs strategy for a regional insurance group in the Midwest. In early 2024 his board asked him to present a view on whether the company should invest in an in-house AI capability or rely on third-party vendors. He spent four weeks pulling analyst reports, tracking funding rounds, and mapping competitor announcements. By the time he finished, two of the vendors he had shortlisted had pivoted their pricing model and one had been acquired. "The market moved faster than my slide deck," he said. That is not a research failure; it is the defining property of the market he was researching. The AI investment landscape needs a different analytical frame from most other technology markets, one built for high velocity and structural uncertainty rather than for stable categories and predictable release cycles.

This lesson gives you the tools to read that market clearly: how it is structured, how capital moves through it, what the competitive dynamics actually are, and how to convert that intelligence into better resource allocation decisions inside your own organisation. The goal is not to predict the market. It is to hold a view that stays useful for longer than the four weeks it took to assemble.

How the AI Market Is Structured

The most useful mental model treats the AI market as three distinct layers, each with different economics, different barriers to entry, and different competitive dynamics. Confusing them is the most common analytical error, because a signal that means one thing at the foundation layer often means the opposite two layers up. Price compression at the base is a threat to the companies experiencing it and a tailwind for everyone building on top of them.

Foundation model providers sit at the base. These organisations train and operate large models at enormous cost, since training a frontier model requires hundreds of millions of dollars in compute alone. That creates a structural barrier: only very well capitalised organisations can compete at this layer at all. The dynamics are winner-take-most, but "most" is not "all", because different models have meaningful performance differences across different tasks, so several providers coexist and buyers with varied workloads have real reasons to use more than one.

Middleware and infrastructure sits in the middle. This layer includes API gateways, vector databases (specialised databases that search by meaning rather than exact match), fine-tuning platforms, observability tooling for AI systems, and orchestration frameworks. Capital requirements are far lower and fragmentation is correspondingly higher, with hundreds of companies competing for overlapping problems. Many will be acquired by the layer above or the layer below, which makes vendor durability a genuine procurement question here in a way it is not elsewhere.

The application and workflow layer is where most enterprise AI value is created and captured. These are products built on top of foundation models: AI-enhanced customer relationship tools, legal document review software, code completion assistants, and vertical solutions for healthcare, finance, and logistics. Competition is intense, and the durable advantage is usually domain expertise plus distribution rather than model quality, because the underlying model capability is available to every competitor on roughly the same terms.

Following Capital Flows

Venture capital investment in AI exceeded $60 billion globally in 2024, with roughly 40 percent concentrated in the United States. The headline number is less informative than the distribution behind it. Foundation model companies absorb a share of that capital wildly out of proportion to their number, and a single 2024 round for one foundation model developer exceeded $6 billion on its own. These mega-rounds are misleading indicators of market health, because they reflect the capital intensity of frontier model training rather than the breadth of the opportunity. Reading them as evidence of a booming market for everyone is a category error.

For most organisations the actionable story sits at the application layer, where a large population of well-funded companies is building specialised tools that solve narrow problems well. That is where the vendors you will actually buy from live, and where the failure rate that matters to your procurement risk is concentrated.

A second pattern is worth tracking more closely than the headline totals: corporate strategic investment is rising faster than pure venture capital. The largest cloud platform providers have committed sums running from $4 billion to $13 billion into leading model developers, which signals that they treat access to model capability as infrastructure rather than as a financial bet. This reshapes the competitive landscape in a way that reaches your organisation directly, because those providers can bundle AI capability into cloud contracts you have already signed, putting sustained pressure on standalone AI vendors who must win on product alone.

Reading Competitive Signals

For strategy purposes, published model benchmarks are noisy. Marketing teams publish selective results, evaluation sets leak into training data, and the differences that show up in a leaderboard often do not survive contact with a specific production workload. A more reliable signal is what customers actually run for high-value production work, which surfaces indirectly through procurement decisions, developer surveys, and the pattern of published case studies rather than through vendor claims.

Price compression is the clearest signal the market gives you. The cost of running a capable large language model (a large AI system that generates text, commonly abbreviated to LLM) has fallen roughly ten times every 18 months since 2022. Capability that cost $0.06 per 1,000 tokens in early 2023 costs under $0.01 in 2025. This commoditisation of inference is steadily shifting economic value away from model providers and toward the applications built on them, which is the single most important structural fact for a buyer to internalise. Watch pricing announcements from the major providers closely, because they telegraph competitive pressure and margin erosion earlier and more honestly than any capability claim.

Talent movement is the third signal, and the slowest to be reported. When senior researchers leave a foundation model developer, where they go is worth noting. New frontier efforts tend to form around talent clusters rather than around funding announcements, and talent movement typically predicts capability competition 12 to 18 months ahead of public product news. This is one of the few signals available to an outside observer that leads rather than lags.

Strategic Positioning Frameworks

Market structure is only useful if it changes a decision. Three framings turn the analysis above into positions your organisation can actually take.

The build, buy, or partner decision

Building foundation model capability is not realistic for 99 percent of organisations, because the capital and talent requirements are prohibitive and the resulting asset depreciates as the frontier moves. The real question sits at the application layer: do you build a custom AI workflow on top of commodity models, buy a vendor solution, or partner with a specialist? The answer turns on whether AI gives you a durable competitive advantage in that specific domain. If it does, build or acquire, and accept the cost of maintaining it. If it is table stakes that every competitor will soon have, buy, and spend the saved effort where you are actually differentiated.

Vendor dependency risk

As AI becomes core infrastructure rather than an experiment, vendor concentration becomes a real operational risk. An organisation whose customer support, document processing, and sales forecasting all run through a single proprietary model interface has one point of failure that is simultaneously contractual, commercial, and technical. A pricing change, a policy change, or an outage propagates through every dependent process at once. Diversification across two or three providers, or maintaining a self-hosted fallback for the most critical path, is ordinary risk management. Treat it the way you treat banking relationships: do not concentrate critical operations with a single counterparty, however good the terms currently are.

The talent market as a leading indicator

Compensation for AI practitioners, meaning the premium organisations pay for people who can implement and operate AI systems rather than research them, has risen 30 to 50 percent in three years for experienced roles. That premium will compress as practical AI skills become widespread, which is exactly what structured practitioner training is designed to accelerate. The strategic implication is that buying scarce specialists is an expensive and temporary answer, while building broad AI fluency across an existing workforce produces an advantage that persists after model costs and salary premiums both fall.

What to Monitor on an Ongoing Basis

Useful market intelligence on AI does not require reading research papers. A focused monitoring practice takes about two hours per month and covers what actually changes decisions. The point of the list below is not comprehensiveness; it is that each item is a leading indicator rather than a summary of what already happened.

• Model release announcements from the major providers, noting capability claims and, more importantly, pricing changes.
• Merger and acquisition activity at the middleware layer, since acquisitions reveal which capabilities larger players consider strategic rather than optional.
• Regulatory developments in the jurisdictions where you operate. The EU AI Act and various US state-level proposals affect procurement criteria and compliance cost directly.
• Developer adoption patterns, including star counts on public code repositories and the results of the large annual developer surveys, which reflect where practitioners are placing their bets well before enterprise procurement follows.
• Competitor job postings. A competitor opening 20 AI engineering roles is announcing a capability build that will affect your competitive position in 12 to 18 months, considerably earlier than any press release would.

Anti-Patterns

• Reading funding headlines as market health. Mega-rounds at the foundation layer measure the capital intensity of frontier training, not the size of the opportunity available to a buyer or an application builder.
• Treating published benchmarks as procurement evidence. Benchmark results are selectively published and frequently fail to predict performance on a specific production workload. Test on your own task.
• Analysing the market as one market. A signal that threatens the foundation layer often benefits the application layer. Applying a single narrative across all three layers produces confidently wrong conclusions.
• Concentrating every critical process on one provider because the current terms are good. Contractual, pricing, and reliability risk arrive together and affect every dependent process simultaneously.
• Building where you should buy. Custom development is justified by durable competitive advantage in a specific domain, not by the fact that the team is capable of building it.
• Solving the talent question by hiring alone. Specialist compensation premiums are a temporary market condition; broad internal fluency outlasts them and is harder for competitors to copy.
• Refreshing the market view only when a decision forces it. Marcus assembled four weeks of research that expired before he presented it. A standing monthly habit produces a view that is current when the question arrives.

Practice Prompts

• List every AI capability your organisation currently pays for and assign each to one of the three layers. Note how much of your spend sits at the application layer, where most enterprise value is created.
• For your most business-critical AI dependency, write down what happens operationally if that provider changes its pricing, its terms, or its availability.
• Find the pricing history for the model interface you use most and calculate what the same workload cost when you first adopted it. Decide what you would do with the difference.
• Take one AI capability you are considering building and argue the opposite case in writing: why buying it would be the better decision, and what would have to be true for that to be wrong.
• Pull the public job postings of two competitors and describe, in a paragraph, what capability each appears to be assembling and when it would begin to affect you.
• Design your two-hour monthly monitoring routine by naming the specific sources you will check for each of the five categories above, then run it once before deciding whether it is worth keeping.

Reflection

Marcus's problem was not the quality of his research. It was that he treated market intelligence as a project with a deadline rather than as a standing practice, so his understanding was at its sharpest on the day it stopped being current. Consider how your own organisation forms its view of the AI market. Is it assembled on demand when a decision is pending, or maintained continuously at low cost? The second approach produces worse individual documents and much better decisions, because the view is already there when someone finally asks.

Glossary

• Foundation model: A large general-purpose model trained at high cost by a well-capitalised organisation and made available to others to build on.
• Middleware layer: The tooling between models and applications, including API gateways, vector databases, tuning platforms, observability, and orchestration frameworks.
• Vector database: A specialised database that searches by meaning rather than exact match, used to give applications access to relevant material at query time.
• Inference cost: The cost of running a trained model to produce an output, usually priced per unit of text processed and falling steadily.
• Price compression: The sustained decline in the cost of equivalent model capability, which shifts economic value from model providers toward application builders.
• Vendor concentration risk: The exposure created when multiple critical processes depend on a single provider, combining contractual, pricing, and reliability risk.
• Leading indicator: A signal that moves before the outcome it predicts, such as talent movement or job postings, as distinct from a report of what already happened.

Related Lessons

• AI Investment & Capital Markets goes deeper into how capital is raised and deployed across the layers described here.
• AI Pricing & Business Models develops the price compression argument into the commercial models buyers actually negotiate.
• Economics of AI & Competitive Advantage examines where value accumulates once model capability itself stops being scarce.
• Industry Analysis & Competitive Dynamics applies structural analysis of this kind to a specific sector rather than to the market as a whole.
• Strategic Positioning & Competitive Advantage takes the build, buy, or partner decision into a full positioning strategy.

Closing

Reading the AI market well is less about forecasting than about knowing which facts are structural and which are noise. Layer economics are structural. Price compression is structural. Talent movement leads. Funding headlines, benchmark tables, and product announcements are mostly weather. Marcus rebuilt his practice around that distinction and stopped trying to produce a definitive answer, which turned out to be the right response to a market that will not hold still long enough for one. When his board asked the same question again, he did not need four weeks.

Key Takeaways

• The AI market has three distinct layers, foundation models, middleware and infrastructure, and applications, with different economics and investment logic at each. Most organisations compete and buy at the application layer.
• Price compression is the key dynamic to track. Inference costs have fallen roughly ten times every 18 months, which shifts economic value from model providers toward application builders and is a structural tailwind for adoption.
• Strategic investment by the largest cloud platform providers reshapes the landscape. Treating model access as infrastructure lets them bundle AI capability into existing cloud contracts and pressures standalone vendors.
• Vendor concentration is a risk worth managing. As AI becomes operational infrastructure, dependence on a single proprietary provider creates contractual, pricing, and reliability exposure at once.
• Talent signals lead product markets. Compensation premiums and competitor job postings are reliable indicators of capability shifts roughly 12 to 18 months ahead of announcements.
• Build, buy, or partner belongs at the application layer. Building foundation model capability is not viable for most organisations; the real question is custom workflows on commodity models versus vendor solutions.
• Maintain the market view continuously. About two hours a month on releases, acquisitions, regulation, developer adoption, and competitor hiring beats four weeks of research that expires on delivery.

Frequently Asked Questions

If inference keeps getting cheaper, should we simply wait? Waiting is a defensible position on capability that is genuinely commoditising, since the same result will cost less later. It is a poor position on everything the price does not cover, which is where the real work sits: understanding your own processes, preparing your data, and building the internal fluency to use the capability when you do adopt it. Those take time that falling prices do not shorten, and organisations that wait on all of it arrive late with the cheap model and none of the readiness.

How many model providers should we actually use? Enough that no single provider's pricing, policy, or availability change can halt a critical process, which for most organisations means keeping a viable second option genuinely tested rather than merely identified. Diversification across two or three providers, or a self-hosted fallback on the most critical path, is proportionate. The failure mode to avoid is a documented alternative that nobody has ever run at production volume.

Our board wants a forecast, not a framework. What do we give them? Give them the structural facts and the decisions those facts change, which is what a forecast is actually being requested for. Layer economics, the direction and rough pace of price compression, and your current concentration exposure support a specific recommendation about where to spend. A point prediction about which provider leads in the long run supports nothing, and it will be wrong in a way the framework would not have been.