Top AI Companies to Invest In: A Strategic Evaluation Framework

Investors keep asking the same question in different ways: which are the top AI companies to invest in, and how do you tell the compounders from the casualties? The honest answer is that any list of "best AI stocks" is a snapshot of a market that is still inventing its own economics. Nvidia, OpenAI, Microsoft, Anthropic, Palantir, TSMC, Databricks, Synthesia — these names appear on every ranking, but their business models, moats and risk profiles are so different that treating them as a single theme is how portfolios blow up.

This guide takes a different approach. Rather than hand you another numbered list, we walk through the full AI value stack, name the companies that genuinely lead each layer, and give you the evaluation framework we use with the venture, private-equity and enterprise clients who commission iCentric Agency to perform technical due diligence on AI businesses. We build and ship AI platforms for a living, which means we see which vendors actually work in production, which demos never survive a procurement conversation, and which "AI revenue" is really just repackaged software maintenance.

What follows is long, deliberately. If you are allocating serious capital — personal portfolio, family office, corporate treasury or venture mandate — a thirty-second ranking is not enough. You need to understand the shape of the market, the choke points, the credible competitors to the obvious incumbents, and the specific risks that could turn today's leaders into tomorrow's AOLs.

Why "investing in AI" needs a framework, not a shopping list

Most listicles about AI companies to invest in share a structural flaw: they conflate three very different kinds of exposure. There is the picks-and-shovels trade, where you own the semiconductor fabs, the networking suppliers and the hyperscalers who sell compute regardless of which model wins. There is the model-layer trade, where you take a view on which foundation model lab will own the next generation of reasoning, multimodal and agentic capability. And there is the application trade, where you bet on specific software businesses that will capture durable margin by embedding AI into vertical workflows.

Each of those trades has a different time horizon, a different sensitivity to interest rates, and a different relationship to the pace of model progress. Owning Nvidia and owning Harvey are not the same bet with a different ticker; they are opposite ends of a value chain. The picks-and-shovels names benefit when everyone is building; the application names benefit when the building stops and workloads consolidate on a few winners. A sensible AI allocation holds some of both, in a ratio tuned to how convinced you are about the pace of commoditisation at each layer.

The second problem with the shopping-list approach is survivorship. The companies that will dominate AI a decade from now are not all public today. Some are three-person labs that have not raised a Series A. Others are divisions of incumbents that have not yet separated out the AI P&L. Any ranking that only looks at listed names with current revenue is systematically biased toward the obvious. That is why serious investors treat the public AI universe as roughly half the opportunity set — the other half sits in private markets, accessible through venture funds, pre-IPO secondaries, and the balance sheets of large-cap tech companies that have become de facto AI index funds.

The third problem is definitional. What counts as an "AI company"? If the test is "generates revenue from AI products", then Microsoft, Alphabet and Meta all qualify, but so does your bank's chatbot vendor. If the test is "would not exist without AI", then OpenAI and Anthropic qualify but Nvidia — a graphics company before CUDA — does not. We prefer a sharper test: a company is an AI investment if the thesis depends on AI adoption being correct. Under that definition, Nvidia is in, Microsoft is partially in (the Azure/Copilot segments only), and your bank's chatbot vendor is firmly out.

This guide is written for allocators who want to make considered, defensible decisions about AI exposure: individual investors with meaningful portfolios, wealth managers constructing thematic sleeves, corporate strategy teams scanning for M&A or partnership targets, and operators who want to understand the competitive landscape they are entering. It is not investment advice. It is a map of the territory, built from the ground truth of actually shipping AI products.

The AI value stack: six layers every investor should map

Before naming any company, map the stack. The AI value chain breaks into six layers, and each layer has a distinct economic signature. Understanding where margin pools, where competition is already fierce, and where moats are forming is the single most important discipline when choosing which AI companies to invest in.

Silicon and accelerators sit at the base. This is the GPU, TPU, NPU and custom-ASIC layer — the physical hardware that runs training and inference. The economics here are classic semiconductor economics: enormous capex, long design cycles, huge gross margins when you have architectural leadership, brutal commoditisation when you do not. The current cycle is unusual because demand has outstripped supply for an extended period, which has let the leader (Nvidia) sustain gross margins that would normally invite far more aggressive competitive response. That gap is closing.

Data-centre infrastructure and energy is the layer most investors underweight. AI workloads are grid-scale consumers of power, water and specialised real estate. The companies building AI-ready data centres, the utilities serving them, the HVAC and liquid-cooling specialists, and the engineering firms siting new capacity are all beneficiaries of the same capex wave that drives Nvidia's order book. This layer is less sexy, less covered, and often better valued.

Hyperscaler cloud platforms aggregate the first two layers and sell them as a service. Microsoft Azure, Google Cloud, Amazon Web Services, Oracle Cloud Infrastructure and a short list of specialists (CoreWeave, Crusoe, Lambda) are the gatekeepers. Hyperscalers capture a substantial share of model-training spend and almost all of the inference spend for enterprises that do not want to run their own hardware. Their AI revenue is high-margin and compounds, but their capex requirements are vast.

Foundation model labs are the names that dominate headlines: OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral, Cohere, xAI, DeepSeek. They train the frontier models that everything else builds on. The economics here are the most contested in the stack. Training costs are rising geometrically, model lifespans are shrinking, and differentiation is thinning as open-weight alternatives narrow the capability gap. Winning here requires either unique data, unique distribution or a cost structure the competition cannot match.

Data and orchestration platforms are the plumbing that makes models useful inside an organisation: Databricks, Snowflake, Palantir, ServiceNow, MongoDB, Elastic, LangChain, LlamaIndex. These companies turn raw models into deployable, governed, auditable systems. This layer is where many enterprise buyers are quietly spending the majority of their AI budget, because without it the models cannot touch real data safely.

Application-layer software and vertical AI sits at the top. Adobe for creative, Salesforce for CRM, SAP for ERP, Harvey for legal, Abridge for clinical documentation, Synthesia for video, ElevenLabs for voice, Decagon and Sierra for customer service. These companies take foundation models and wrap them in workflow, data and distribution. If model layers commoditise, this is where the durable margin accrues.

With that map in hand, we can now look at who leads each layer and why.

Layer 1 — Chipmakers and semiconductor leaders

Nvidia

Nvidia is the default AI investment and will remain so until something changes materially. The company has three overlapping moats: CUDA (a software ecosystem that has had more than fifteen years of head start), NVLink and networking (the Mellanox acquisition made Nvidia a systems company, not a chip company), and manufacturing allocation at TSMC that competitors struggle to match. Gross margins are extraordinary and have held up through multiple product cycles.

The risks are not what most people assume. The near-term risk is not that AMD or custom silicon takes meaningful share — that is a slow-motion story. The near-term risk is customer concentration. A short list of hyperscalers accounts for a very large share of Nvidia's data-centre revenue. If one of them pauses capex, revises its depreciation schedule, or accelerates its own ASIC programme, the earnings sensitivity is severe. Investors buying Nvidia today should model what happens when hyperscaler capex growth normalises, not when it stops.

The longer-term risk is architectural. Transformer-dominant workloads suit Nvidia's GPUs well. If reasoning, agentic and mixture-of-experts architectures shift the inference economics toward specialised silicon — which several labs are now exploring — the moat erodes. Nvidia is aware of this and is iterating product cycles aggressively, but the risk is real.

Advanced Micro Devices (AMD)

AMD is the most credible second source. The MI-series accelerators have closed much of the raw-performance gap, ROCm has matured enough that major labs will deploy on it for inference, and the acquisition of ZT Systems gave AMD rack-scale integration capability. Hyperscalers want a credible alternative to Nvidia for pricing leverage alone, and AMD is the only Western option at scale.

The investment case rests on two things: whether AMD can grow its data-centre GPU revenue to a meaningful share of the market, and whether the server CPU franchise (EPYC) continues to take share from Intel. Both have been playing out. Gross margins remain below Nvidia's and probably always will, but the valuation discount reflects that. For investors who believe the AI accelerator market will be a two-player market long-term, AMD is the obvious complement to Nvidia.

Broadcom

Broadcom is a stealth AI play. The custom-silicon business designs ASICs for the hyperscalers who want their own training and inference chips — Google's TPUs are the best-known example, but there are others. Add the networking franchise (Tomahawk, Jericho switches that power the fabric inside AI clusters) and Broadcom captures AI spend that bypasses Nvidia entirely. The VMware acquisition broadened the software franchise and gave Broadcom a durable enterprise distribution footprint.

The risk is executional. Hyperscaler ASIC programmes are lumpy, with design wins that take years to materialise and can be cancelled. The VMware integration has been contentious with customers. Broadcom's capital allocation under its current management has been excellent, but the complexity of the business makes it harder to analyse than a pure-play.

Taiwan Semiconductor Manufacturing (TSMC)

TSMC is the most important company in AI that most retail investors underweight. Every leading-edge accelerator — Nvidia, AMD, the hyperscaler ASICs, Apple's silicon, Qualcomm's AI chips — is manufactured by TSMC. There is no credible second source for leading-edge logic at the volumes required. Samsung Foundry is a distant second; Intel Foundry Services is an ambition rather than a reality.

The risk is geopolitical and existential. The Taiwan Strait is the single greatest concentration risk in the global technology supply chain. Investors buying TSMC are being paid a premium for taking that risk, and the premium has arguably compressed as capacity expansion in Arizona, Japan and Germany makes the supply chain marginally more resilient. For most portfolios, TSMC is a core AI holding; sizing it is the harder question.

Intel

Intel is a turnaround trade, not an AI pure-play. The Gaudi accelerator line has not gained meaningful share. The foundry business is a credible long-term strategic asset but a near-term capital sink. Buyers of Intel are betting that government subsidies, a successful Intel 18A process node, and a disciplined refocus on manufacturing restore the franchise. That bet may pay off, but it is a semiconductor bet, not an AI bet, and should be sized accordingly.

ARM, Marvell and the supporting cast

ARM Holdings earns a royalty on an enormous share of the compute shipped into AI endpoints and increasingly into data-centre CPUs (Nvidia Grace, AWS Graviton). Marvell is a leader in optical networking and custom silicon for hyperscalers. Micron, Samsung and SK Hynix dominate high-bandwidth memory (HBM), which is a genuine bottleneck in modern accelerators and often undervalued relative to its strategic importance. ASML, the sole supplier of EUV lithography equipment, is the chokehold above the chokehold.

Layer 2 — Hyperscalers and cloud platforms

Microsoft

Microsoft is the hyperscaler with the most consequential AI partnership (OpenAI), the most successful AI consumer product (Copilot in Office), and the most defensible enterprise distribution (the installed base of Windows, Azure Active Directory, and the Microsoft 365 suite). The thesis is straightforward: AI accelerates seat-level monetisation of software that was already in every enterprise, and Azure captures the compute spend of the companies that train and run AI models.

The questions are two. First, how durable is the OpenAI relationship as OpenAI's own corporate structure evolves and as it seeks additional compute suppliers? Second, how much of Copilot's attach-rate reflects genuine productivity value versus procurement bundling? The evidence from our own enterprise clients suggests that Copilot adoption is real but uneven; the companies that have redesigned workflows around it see large gains, while those that have treated it as a feature toggle see little. Microsoft's challenge is to drive the former pattern.

Azure's AI revenue has grown faster than any other hyperscaler segment. The capex required to sustain that growth is enormous. Investors who are comfortable with the capex intensity and believe enterprise AI adoption continues to broaden have a straightforward case for Microsoft as a core AI holding.

Alphabet

Alphabet is the hyperscaler with the most complicated AI story. The upside is that Google has unique assets — proprietary silicon (TPUs), a world-class research lab (DeepMind), the deepest multimodal data in the industry (YouTube, Search, Maps), and a credible frontier model (Gemini). The downside is that AI threatens the search advertising business that generates most of the group's cash flow. If users increasingly satisfy information needs through conversational agents rather than ten blue links, click volumes compress and so does revenue per query.

Google's response has been to integrate AI into search (AI Overviews), to build Gemini into the Workspace suite, and to push TPUs as a differentiated cloud offering. The strategic logic is sound. The question is whether the pace of search monetisation decline exceeds the pace of new AI revenue growth during the transition. For investors with a longer time horizon, the asset base justifies patience. For those focused on near-term earnings, the ambiguity is uncomfortable.

Amazon

Amazon's AI story is the quietest of the three giants. AWS Bedrock gives enterprise customers access to multiple foundation models (Claude from Anthropic, Llama from Meta, Mistral, Amazon's own Nova family) through a single API, which fits the "model-agnostic enterprise" pattern we see across our client base. Trainium and Inferentia are credible custom-silicon alternatives for workloads where cost per token matters more than raw performance. Anthropic is a significant strategic partner.

The investment case is that AWS continues to be the default cloud for enterprises that want optionality, that Amazon's own retail and logistics operations provide a vast internal AI training ground, and that advertising continues to grow as a secondary beneficiary of AI-driven targeting improvements. The risk is that AWS's growth rate decelerates faster than Azure's as hyperscaler share shifts around AI workloads.

Oracle

Oracle has been one of the surprise beneficiaries of the AI capex cycle. OCI (Oracle Cloud Infrastructure) was a laggard for most of the cloud era, but its architectural choices — particularly its RDMA networking fabric — turn out to suit large-scale GPU clusters well. Oracle has signed meaningful AI compute contracts and is building out capacity aggressively. The database franchise provides a stable earnings base and a natural AI data-gravity story.

The risk is the backlog-to-revenue conversion ratio. Oracle has disclosed very large contracted amounts, but revenue recognition depends on building and populating the data centres that will serve those contracts. The capex is real, the counterparty concentration is high (a small number of very large AI customers), and the execution risk is non-trivial. For investors willing to underwrite the delivery, Oracle offers more upside than the mega-cap hyperscalers at a lower multiple.

Meta

Meta is best understood as an AI-advantaged advertising business with an optional foundation-model upside. The Llama family of open-weight models has shaped the open-source landscape, but the direct monetisation of those models is modest. The real AI contribution to Meta's P&L comes from ranking, targeting and content moderation improvements that lift ad revenue per user. Those gains have been significant and are likely to continue.

Meta's capex intensity has risen sharply as it builds AI infrastructure, which has unsettled some investors. The counter-argument is that the ad business generates enough free cash flow to absorb the investment without balance-sheet strain, and that each percentage point of ad-revenue uplift from better AI pays for a lot of GPUs. Reality Labs remains a loss centre and should not be in the thesis.

Weighing capex-heavy giants against pure-plays

The mega-cap hyperscalers trade at modest multiples relative to the AI pure-plays, but their AI exposure is diluted by legacy businesses. A rough rule: if you want direct AI beta, you need the chipmakers and the specialists; if you want AI exposure with downside protection from mature cash-generative businesses, the hyperscalers do that job. Most sensible AI allocations hold both.

Layer 3 — Foundation model labs and private AI pure-plays

OpenAI

OpenAI is the most consequential private company in the AI stack. ChatGPT redefined consumer software distribution, the API has become the default way enterprises experiment with large language models, and the product cadence (GPT, DALL-E, Sora, Codex, operator-style agents) has consistently set the industry agenda. Revenue growth has been exceptional.

The investment questions are structural rather than product-related. OpenAI's corporate structure — a capped-profit entity under a non-profit — has evolved through multiple governance iterations, and the direction of travel is toward a more conventional for-profit form that would enable an eventual public listing. The partnership with Microsoft is central to the compute story but creates obvious strategic tension as OpenAI diversifies its infrastructure footprint. Costs are vast; profitability is a question of when, not whether, model costs decline faster than enterprise prices.

Retail investors cannot buy OpenAI directly. Exposure runs through Microsoft, through secondary platforms that trade in late-stage private shares, and through venture funds that have participated in previous rounds.

Anthropic

Anthropic is the credible frontier competitor to OpenAI. Claude has emerged as the model of choice for a particular enterprise segment — organisations that value safety research, long-context performance, and the Constitutional AI approach to alignment. Claude Code has made meaningful inroads in developer workflows. The partnerships with Amazon (as a strategic investor and primary cloud) and Google (as a secondary cloud and investor) give Anthropic infrastructure optionality that OpenAI's Microsoft-centric arrangement lacks.

The thesis is that in a market with two or three frontier labs, being clearly number two is a very good business. The risk is the same as OpenAI's: training costs rise faster than revenue, and open-weight competitors erode pricing power at the mid-market.

Mistral

Mistral is the European champion. Headquartered in Paris, with open-weight models that have been adopted broadly and a growing enterprise business, Mistral has become the default choice for organisations that want foundation-model capability without exclusive dependence on US providers. European government agencies, in particular, have gravitated to Mistral for sovereignty reasons.

For investors, Mistral offers something no US lab does: a non-US frontier model developer with a credible path to continued independence. The valuation reflects scarcity. Access is private-market only.

Cohere

Cohere has taken a different path. Rather than competing for consumer attention, it has focused on enterprise retrieval, private-cloud deployments and verticals where data cannot leave a customer's environment. The partnership with Oracle and the UK government's engagement with Cohere have given it a profile beyond its headline model performance.

The investment case is that enterprise foundation-model demand will skew heavily toward deployments that respect data residency, security and compliance requirements that consumer-oriented models cannot easily satisfy. Cohere is best positioned for that segment.

xAI, Safe Superintelligence, Thinking Machines Lab and the talent trade

A cluster of newer labs are better understood as talent-concentration vehicles than as product companies. xAI has access to Twitter/X data and to Elon Musk's capital allocation. Safe Superintelligence (Ilya Sutskever) has raised substantial amounts without a product. Thinking Machines Lab (Mira Murati) and Reflection are following similar paths. Physical Intelligence and Skild AI are building foundation models for robotics.

These are venture-style bets with venture-style outcomes. The expected value could be very high; the probability of individual success is modest. Exposure is private-market only and is best taken through diversified venture funds rather than concentrated secondary positions.

How retail investors reach private AI labs

The practical options are: funds-of-funds with venture AI exposure, pre-IPO secondary platforms (with the usual illiquidity, information asymmetry and fee drag), listed holding companies and investment trusts with AI-weighted portfolios, and large-cap tech stocks that have taken stakes in the labs (Microsoft for OpenAI, Amazon and Google for Anthropic, Nvidia across many). Each route has trade-offs; the last is the cleanest for most individual investors.

Layer 4 — Enterprise data and AI platforms

Databricks

Databricks sits at the intersection of data engineering, machine learning and now foundation-model serving. The lakehouse architecture — unifying data warehouse and data lake patterns — has become the default for new data platforms at scale. The Mosaic AI acquisition brought model-training capability in-house, and the company's open-source commitments (Delta Lake, MLflow, Unity Catalog) have cemented developer mindshare.

For investors, Databricks is a pre-IPO holding available through secondary markets and through venture vehicles. The business is one of the clearest beneficiaries of enterprise AI budgets: when a bank or retailer builds an AI platform, Databricks is almost always on the shortlist.

Snowflake

Snowflake went public earlier and has had a more complicated journey. The core data-warehouse franchise is excellent, and the data-sharing architecture is a genuine moat. The AI story (Cortex, Snowpark, native model serving) is credible but has taken longer to monetise than investors initially expected.

The thesis for Snowflake rests on data gravity: once an enterprise has consolidated its data in Snowflake, it is natural to run AI workloads where the data already lives. The counter-thesis is that Databricks is competing hard for exactly that workload and that open table formats (Iceberg, Delta) reduce the switching costs that previously protected Snowflake's position.

Palantir

Palantir has become one of the most polarising names in AI. The bull case is that AIP (the Artificial Intelligence Platform) has found product-market fit with large commercial enterprises that previously found Palantir inaccessible, that government revenue remains a stable backbone, and that the ontology-first approach to enterprise data is uniquely suited to agentic workloads. The bear case is valuation, concentration risk and a culture that is difficult for traditional institutional investors to underwrite.

From the technical perspective we take on evaluation projects, AIP is a genuinely differentiated product for a specific use case: organisations that need to connect multiple data systems and apply AI to operational decisions with audit trails. For that use case, few competitors come close. The valuation question is separate.

ServiceNow

ServiceNow has quietly become one of the most important enterprise AI platforms. The Now Assist suite embeds generative AI into IT service management, HR workflows, customer service and increasingly broader business process workflows. The incumbency inside large enterprises is extensive, the data model is well-suited to agentic automation, and the pricing structure has shifted toward consumption in a way that captures AI-driven usage expansion.

For investors looking for AI exposure through a cash-generative, enterprise-anchored business with established distribution, ServiceNow is one of the cleanest expressions available in public markets.

MongoDB, Elastic and the database race

The database layer is being reshaped by vector search, hybrid retrieval and the need to combine operational and analytical workloads. MongoDB has added native vector search to its document database. Elastic has strong positioning in retrieval. Dedicated vector databases (Pinecone, Weaviate, Qdrant) are competing with incumbents who are rapidly adding vector capability. For investors, the question is whether specialised vector databases retain enough differentiation to build durable businesses, or whether vectors become a feature of every mainstream database. Our view, based on what we see in production systems, is that vectors alone are losing their standalone case; combined operational-plus-vector databases are winning.

Why picks-and-shovels may outlast the model wars

If frontier model performance converges (and it is converging faster than most consumer narratives acknowledge), the differentiated economics shift to the layer that governs, routes, retrieves and monitors models in production. That is the Layer 4 opportunity. Investors who want AI exposure that is less exposed to model-layer commoditisation should overweight this layer relative to the model labs.

Layer 5 — Application-layer leaders and vertical AI

Adobe

Adobe is the clearest example of an incumbent that AI could either make or break. The Firefly model family and the integration of generative capability across Photoshop, Illustrator, Premiere and Express are genuine. The commercial-safe training set (Adobe Stock) addresses the enterprise concern about generative content and copyright. The Creative Cloud distribution is unmatched.

The risk is that creative AI tools from newer entrants (Midjourney, Krea, Runway, Ideogram) and from large tech platforms erode Adobe's monopoly on professional creative workflows. The counter-argument is that enterprise buyers want the governance, licensing and integration Adobe provides, and that the professional workflow lock-in is deeper than consumer-grade competitors appreciate.

Salesforce

Salesforce has pivoted hard to Agentforce, its agentic AI platform built on Data Cloud. The pricing is consumption-based, which aligns incentives with usage and gives Salesforce upside as agents take on more work. The installed base of CRM, marketing and service cloud customers is a vast distribution advantage.

The execution risk is real. Agentic AI deployments require more customer services effort than traditional SaaS implementations, and the gross margin profile at scale is still being established. Salesforce's historical capital allocation has been uneven. For investors who believe that enterprise agents will consolidate on a small number of platforms with strong CRM data gravity, Salesforce is a plausible long. For those who believe agentic workloads will fragment across many vendors, the thesis is weaker.

SAP

SAP is the enterprise AI story that European investors underweight. The Joule assistant, Business AI across the S/4HANA footprint, and the Datasphere integration with Databricks give SAP a credible AI stack built on the operational data that runs much of global commerce. The move to cloud ERP has been slow but is now compounding. For investors who want AI exposure tied to the hardest-to-displace layer of enterprise software, SAP merits attention.

Synthesia, ElevenLabs and the UK/European media stack

Synthesia (London) is the leader in enterprise AI video generation. ElevenLabs (UK/US) dominates voice. These companies have built genuine enterprise businesses with meaningful recurring revenue and distinguish themselves from consumer-media tools through governance, voice cloning controls and integration with corporate learning and communications platforms. Both are private; exposure is venture-only but the UK presence makes them relevant for domestic investors looking at the ecosystem.

Harvey, Abridge, Decagon and the vertical agent playbook

The vertical AI thesis is that general-purpose models commoditise and durable margin accrues to companies that combine model capability with proprietary workflow, data and distribution in a specific industry. Harvey (legal), Abridge (clinical documentation), Decagon and Sierra (customer service), OpenEvidence (medical reference), Rogo (investment banking), Clay (sales operations) and EliseAI (property management) are examples. Each has achieved meaningful enterprise penetration in a narrow domain.

These are the companies most likely to produce the next generation of enterprise software winners. They are also the most likely to be acquired by incumbents before reaching public markets. Exposure is predominantly private.

Spotting AI revenue vs AI-washing

Investor relations departments have become fluent in AI narrative. The useful tests are: does the company disclose AI-attributable revenue separately; is AI gross margin disclosed and is it holding up; are AI customers also paying customers of the core product (expansion) or new customers (acquisition); is the capex-to-AI-revenue ratio sensible. Companies that cannot answer these questions clearly are usually AI-adjacent rather than AI-native.

Layer 6 — Robotics, physical AI and the next decade

The current generation of AI investments is dominated by digital workloads. The next generation will extend into the physical world. Physical Intelligence and Skild AI are building foundation models for robots using large-scale teleoperation data. Figure, 1X, Agility and Apptronik are building humanoid platforms. Boston Dynamics remains the technical benchmark. Tesla is simultaneously a vehicle company, a battery company and a robotics company depending on which analyst you ask.

Industrial automation incumbents — Siemens, Rockwell Automation, ABB, Fanuc, Keyence — are the obvious listed beneficiaries of AI capability moving into factories and warehouses. Their AI revenue is still a small fraction of the total, but the installed base is enormous and the integration advantage is real.

Defence-adjacent AI deserves separate consideration. Palantir, Anduril, Shield AI and prime contractors with AI programmes (Lockheed Martin, Northrop Grumman, Leonardo, BAE Systems) are a distinct allocation decision with its own ethical and regulatory considerations. For investors who are comfortable with the sector, defence AI budgets are durable and growing.

The physical AI thesis is a longer-duration bet. The technology is less mature, the deployment cycles are longer, and the capital intensity is higher. For portfolios with multi-decade horizons, a modest allocation is defensible; for shorter horizons, the opportunity may be too early.

International picks worth a serious look

AI is not a US-only investment theme, though the US dominates. For UK and European investors, international diversification within the AI theme can reduce concentration risk and capture value that US-centric lists miss.

TSMC and ASML — the non-negotiable supply chain

We covered TSMC above; ASML belongs in any AI allocation. The Dutch lithography company is the sole supplier of EUV equipment and the dominant supplier of DUV. Every leading-edge chip produced anywhere in the world depends on ASML tooling. The order book is long, the margins are strong, and the moat is as deep as any in global technology.

Tencent and Alibaba — Chinese AI exposure

Chinese AI has advanced faster than Western investors generally appreciate. DeepSeek's cost-efficient model training rattled markets and demonstrated that frontier capability does not require the capex intensity of the leading US labs. Tencent and Alibaba have both built credible foundation-model families (Hunyuan and Qwen respectively), both operate hyperscale cloud businesses, and both have unique data assets in Chinese consumer and commercial markets.

The investment considerations are policy and sentiment. Chinese tech stocks have traded at discounts to Western peers for reasons that include regulatory uncertainty, geopolitical tension, variable interest entity structures and ESG concerns. For investors willing to size positions carefully, the risk-adjusted upside can be attractive.

Samsung and SK Hynix — the memory bottleneck

High-bandwidth memory (HBM) is a genuine constraint on accelerator throughput and sits in a near-duopoly between SK Hynix and Samsung, with Micron closing in. For investors looking for AI exposure through the component layer, HBM leaders offer a cleaner play than the broader memory cycle might suggest.

Mistral, Synthesia, DeepL — European AI with export potential

Mistral and Synthesia we covered above. DeepL is the German machine translation specialist that has built a durable business against free alternatives from Google and Microsoft by consistently delivering higher-quality translations in professional and legal settings. European AI champions remain under-represented in global AI indices, which creates opportunity for allocators willing to construct positions deliberately.

Currency, listing and tax considerations

UK investors should think carefully about where AI companies are listed and how currency exposure interacts with the position. ADRs, direct foreign-listed exposure through platforms that support it, UCITS-wrapped ETFs and OEICs each have different cost, custody and withholding-tax profiles. A sensible AI sleeve typically mixes several routes.

ETFs, trusts and funds for diversified AI exposure

Single-stock picking in AI is harder than it looks. The leaders change, the valuations are volatile, and the technical judgements required to distinguish real from hyped revenue are specialist. For many investors, diversified exposure through funds is the sensible route.

Thematic ETFs

The thematic AI ETF universe has expanded quickly. iShares Future AI & Tech ETF, WisdomTree Artificial Intelligence UCITS ETF, L&G Artificial Intelligence UCITS ETF and several others compete for the thematic allocation. Investors should look carefully at construction methodology (market-cap weighted vs equal weighted vs thematic-score weighted), overlap with their existing holdings (if you own Microsoft, Nvidia and Alphabet directly, a market-cap AI ETF adds concentration rather than diversification), and total expense ratio.

Semi ETFs vs AI-labelled ETFs

A semiconductor ETF (VanEck Semiconductor, iShares Semiconductor) often captures more AI beta than an AI-labelled ETF because the chip supply chain is where AI capex concentrates. For investors who want pure picks-and-shovels exposure, semi ETFs are frequently the better instrument.

Investment trusts with AI exposure

UK investors have the advantage of a mature investment-trust market. Scottish Mortgage, Allianz Technology Trust, Polar Capital Technology Trust and Manchester & London Investment Trust all have meaningful AI exposure, with the added benefit of discount dynamics that can create entry opportunities. Trusts also enable private-company exposure that open-ended funds cannot easily provide.

Combining active and passive

A sensible AI allocation often combines a passive thematic sleeve (for broad exposure at low cost) with an actively managed trust or fund (for private-company and judgement-based exposure that passive products cannot provide). The ratio depends on the investor's conviction in active management for this specific theme.

A sample allocation framework

For an investor with meaningful AI conviction and a long time horizon, a plausible AI sleeve might look like: a core allocation to semiconductor and hyperscaler exposure (mixing direct holdings in two or three mega-caps with a semi ETF), a satellite allocation to Layer 4 enterprise platforms (Palantir, ServiceNow, Databricks via a trust that holds it), a smaller allocation to vertical application leaders, and a venture-style sleeve for private market exposure. Position sizes should respect the volatility of the theme; a sleeve that is comfortable in a bull market is often badly sized for a drawdown.

Risks and red flags when buying AI stocks

Circular revenue and vendor financing

One of the most important dynamics to watch is the circularity of AI revenue. Hyperscalers invest in model labs; the labs spend the capital on compute from the same hyperscalers; the hyperscalers book the compute as revenue. Nvidia invests in cloud providers and model labs; those companies buy more Nvidia hardware. None of these flows are necessarily inappropriate, but they create a cycle in which the strength of the market depends on continued capital commitment from a small number of players. If any link slows, the chain adjusts rapidly.

Capex intensity and the depreciation cliff

AI infrastructure has useful lives that are shorter than previous generations of data-centre hardware. Depreciation schedules have been extended at several hyperscalers in ways that boost near-term earnings but defer the cost. When the schedules normalise or when workloads shift to require re-equipped capacity, the earnings impact can be material. Investors should model the sensitivity to depreciation assumptions explicitly.

Model commoditisation

Open-weight models are catching up with frontier closed models faster than the leading labs' pricing strategies anticipated. If the gap closes further, the economic case for paying premium rates for frontier model access weakens. Model labs that cannot build sticky distribution beyond raw capability are most exposed.

Regulation

The EU AI Act, the UK's evolving AI governance regime, US export controls on advanced semiconductors and model weights, and sector-specific regulation in healthcare, financial services and employment all shape the AI investment landscape. Companies with weak compliance posture may face material restrictions on deployment, procurement or export. We cover this in depth in our governance and compliance work with enterprise clients.

Energy, grid and water constraints

AI data centres are grid-scale consumers of power and water. In several geographies, grid connection queues are measured in years. Water-stressed regions face real constraints on cooling. The companies that are building AI capacity in power-abundant jurisdictions and investing in behind-the-meter generation will have a structural advantage.

Hype cycles and insider selling

Investor behaviour at AI peaks has rhymed with previous technology cycles. Insider selling calendars, lock-up expirations after IPOs and secondary offerings often mark local tops. Capital-intensive themes with narrative-driven multiples are particularly vulnerable to sentiment-driven drawdowns. Risk management discipline matters more in thematic exposure than in broad market exposure.

A due diligence checklist adapted from enterprise AI buyers

The questions we ask when evaluating an AI vendor for an enterprise client are the same questions investors should ask when evaluating an AI company for a portfolio.

Does the company own proprietary data or just a model wrapper? Many "AI companies" are thin wrappers over foundation model APIs. The valuable businesses have proprietary data, proprietary workflow and proprietary distribution. A wrapper is not a moat.

What is the gross margin profile at steady state? AI gross margins are often depressed in early stages by model inference costs. The question is where margins settle when inference costs decline and when the company has scale. Businesses that cannot articulate a credible steady-state margin profile are red flags.

How defensible is the distribution channel? Enterprise AI distribution runs through hyperscalers, systems integrators, independent software vendors and direct sales. Companies that own their distribution are better positioned than those who depend on partners who could become competitors.

What happens if frontier model costs drop by ninety percent? Model costs have declined dramatically and will continue to decline. If a company's unit economics depend on current model costs, the trajectory is dangerous. If the company benefits from lower model costs (because volume expands faster than per-unit prices decline), the trajectory is favourable.

How exposed is the business to a single hyperscaler? Being hosted on one hyperscaler is operational dependence; depending on one hyperscaler for distribution, co-selling and procurement is strategic dependence. The latter is a material risk that is often under-disclosed.

Is management disclosing AI revenue honestly? The companies that break out AI-attributable revenue, discuss its gross margin and discuss its growth separately are the ones taking the discipline seriously. The companies that bundle AI into legacy segments are either not confident in the numbers or are hiding something.

Portfolio construction: building an AI sleeve that survives a drawdown

Core, satellite and venture buckets

A well-constructed AI sleeve typically has three buckets. The core bucket (perhaps half to two-thirds of the sleeve) holds the picks-and-shovels names that benefit from any version of the AI thesis playing out — semiconductors, hyperscalers, data platforms. The satellite bucket holds specialist names where the investor has specific conviction — a vertical AI leader, a hardware specialist, an international pick. The venture bucket holds private-company exposure through funds or trusts.

Rebalancing rules

Thematic sleeves run hot in bull phases and cold in corrections. Rules-based rebalancing — trimming when a position exceeds a threshold, adding when it falls below — removes emotional decision-making at exactly the moments when emotion is most expensive. Many of the best thematic investors we see use mechanical rebalancing bands rather than discretionary judgements.

Hedging concentration

AI exposure is correlated with the broader large-cap tech trade. Investors who already have significant tech exposure through index funds should think about whether an AI sleeve genuinely adds diversification or simply doubles down. Pairing AI exposure with cash-generative incumbents outside the theme (consumer staples, healthcare, infrastructure) is a useful ballast.

Position sizing for private exposure

Private-market AI exposure is illiquid, information-poor and often marked at stale valuations. Position sizes should reflect that. A rough discipline: do not hold more in private AI than you would be comfortable losing entirely over a five-year horizon.

Tax wrappers for UK investors

UK investors should use ISAs and SIPPs for AI exposure where possible. The volatility of the theme and the potential for significant capital gains make tax-sheltered wrappers particularly valuable. VCT and EIS wrappers can give access to earlier-stage AI companies with meaningful tax benefits, though the risk profile is correspondingly higher.

When to trim

Valuation signals that have worked historically in technology themes include forward price-to-sales multiples deviating significantly from long-run averages, insider selling clusters, the pace of new IPO filings and the behaviour of the IPO-aftermarket performance. None of these are perfect signals, but combinations of them have correlated with local tops in previous cycles.

How iCentric Agency helps investors and operators evaluate AI companies

iCentric Agency builds and ships AI platforms for enterprise clients. That operator perspective turns out to be unusually useful for investment work. Our engagements with investors and corporate strategy teams typically take one of four forms.

Technical due diligence for VC and PE. When a fund is evaluating an AI target, we assess the technology stack, the model and data architecture, the IP position, the engineering practices and the realistic cost structure at scale. We frequently identify issues that commercial due diligence misses — brittle fine-tuning pipelines, undocumented reliance on a single model provider, retrieval architectures that will not scale, governance gaps that will fail enterprise procurement.

Build-vs-buy assessments for enterprises. Many of our clients are evaluating whether to adopt a specific AI vendor, build equivalent capability in-house, or combine both. We produce side-by-side comparisons grounded in real deployment experience, not vendor marketing. These assessments regularly change procurement decisions.

Governance and compliance reviews. We align AI deployments to ISO/IEC 42001, the EU AI Act, UK regulatory expectations and sector-specific requirements. For investors assessing an AI target's regulatory exposure, this work is often the difference between a confident bid and a walk-away.

Case studies from our own delivery. Our published case studies cover AI-powered SEO platforms, agentic content management, carrier rate intelligence, augmented-reality configurators, hotel rate intelligence and reverse logistics. Each illustrates a different pattern of AI value creation and the engineering choices that make it work.

If you are evaluating a specific AI company — public or private — and want an operator's view on the technology, the moat and the risk profile, that is work we do regularly. Brief us with the target, the thesis and the timeframe, and we will scope a sprint that gives you a defensible view before you commit capital.

The short answer to "what are the top AI companies to invest in" is that the question is incomplete without a layer, a timeframe and a risk budget. The longer answer is above. We hope it helps.