Most conversations about AI competition focus on the model leaderboard. OpenAI versus Anthropic. Claude versus ChatGPT. Gemini versus Grok. Which model wins on coding this month. Which one has the best reasoning benchmark. Which demo looked most magical on X.
Models are the part of AI most people can see and touch. They are also the part journalists, investors, customers, and executives can most easily compare.
But looking only at models is like judging the airline industry by the comfort of the seat without asking who owns the planes, the engines, the airport gates, the fuel contracts, the maintenance systems, and the customer relationship.
AI is not one market anymore. It is a stack.
And the companies that win will not simply be the ones with the best model on a given benchmark. They will be the companies that control enough of the stack to shape cost, speed, supply, distribution, and adoption.
That is a very different way to look at the AI competitive landscape.
Why the AI Stack Matters
A year ago, you could almost get away with describing the AI race as a fight among model labs. Today, that view is badly incomplete. The real AI race now stretches across at least six layers:
- Energy, power, and cooling
- Chips, memory, and networking
- Compute capacity and AI cloud
- Frontier models
- Data, developer tools, and model operations
- Apps, distribution, and devices
Each layer creates leverage. Each layer creates bottlenecks. Each layer changes who has power.
A company can lead in one layer and still be dependent in another. A model lab may have brilliant researchers but still need access to scarce compute. A cloud provider may have enormous infrastructure but weaker consumer distribution. A device company may lack public cloud scale but still own the most intimate user interface in the world.
This is why the AI race is becoming less like a software market and more like a systems market.
Think of it like building a city. The model is the beautiful skyline. That is what everyone photographs. But the city only works because of the power grid, roads, water systems, zoning, fiber, ports, permits, construction materials, and daily logistics.
AI has the same problem. The magic at the top depends on hard, physical constraints underneath.
Who Leads Each Layer of the AI Stack
The AI stack is no longer just about frontier models. Leadership is distributed across energy, chips, memory, compute, models, tools, and distribution.
The first image maps the stack layer by layer. The important shift is that the bottom layers now matter as much as the top.
Layer 0: Energy, power, and cooling
This is the layer that used to be invisible in AI strategy conversations. It cannot be invisible anymore.
AI infrastructure is power-hungry. Large-scale training and inference require not only chips, but also electricity, cooling, land, substations, grid access, and the ability to stand up massive data centers quickly.
That is why energy and infrastructure players deserve a place in the AI stack.
Companies like Constellation, NextEra Energy, GE Vernova, Schneider Electric, Eaton, Vertiv, and Siemens Energy are not model companies. They are not building consumer AI assistants. But they increasingly influence who can scale AI at the physical level.
A model company can have the best algorithm in the world. If it cannot get enough power, cooling, and data center capacity, it cannot compete at the frontier for long.
The analogy is simple: AI compute is the factory. Power is the fuel. Cooling is the ventilation. Grid access is the land permit. Without those, the factory does not run.
This is why energy is no longer a background utility. It is becoming a strategic AI bottleneck.
Layer 1: Chips, memory, and networking
This is the most important lower-stack layer.
Everyone talks about GPUs, but the semiconductor layer is broader than NVIDIA alone.
NVIDIA remains the dominant AI accelerator company. It controls one of the most valuable choke points in the entire AI economy. But the rest of the semiconductor ecosystem matters too.
TSMC is the manufacturing backbone. If advanced chips cannot be fabricated at scale, the rest of the stack slows down.
SK hynix, Micron, and Samsung matter because high-bandwidth memory, or HBM, has become a gating resource. The AI model may get the headline, but memory often determines whether the system can actually run efficiently at scale.
ASML matters because advanced lithography is one of the hardest bottlenecks in semiconductor manufacturing.
Broadcom matters in networking, connectivity, custom accelerators, and the plumbing that allows massive AI clusters to work.
Qualcomm matters most at the edge, especially where AI moves into phones, PCs, cars, robotics, industrial systems, and connected devices.
Cerebras and Groq are specialist challengers. Cerebras is focused on wafer-scale AI systems. Groq is focused on fast inference through its LPU architecture.
This layer is where the AI race gets brutally physical.
It is not enough to have a clever model. You need accelerators, memory, packaging, interconnects, networking, and supply chain access. The labs at the top are constrained by the silicon beneath them.
Layer 2: Compute capacity and AI cloud
This is where the hyperscalers come in.
The leaders here are Microsoft Azure, AWS, and Google Cloud. They have the cloud infrastructure, customer relationships, engineering depth, and capital to build massive AI systems.
But this layer is changing fast.
Oracle OCI has become surprisingly relevant because of its AI infrastructure push. CoreWeave has carved out a strong position as a purpose-built AI cloud. And xAI / SpaceXAI is becoming more interesting because of Colossus and its large-scale compute ambitions.
This is where the latest Anthropic compute partnership matters.
The point is not simply that Anthropic got more compute. The bigger point is that compute itself is becoming a strategic product.
When a frontier model company relies on another model-adjacent company’s supercomputing capacity, the old boundaries blur. xAI / SpaceXAI is not just participating in the model race. It is moving into the compute-capacity race.
That matters.
It suggests a future where compute providers are not just neutral infrastructure vendors. They may become strategic gatekeepers, partners, and competitors all at once.
Layer 3: Frontier model builders
This is the layer most people watch.
The leaders remain OpenAI, Anthropic, and Google DeepMind, with strong challengers including Meta, xAI, Mistral, and increasingly model efforts from Amazon and Microsoft.
This layer moves the fastest. Model quality can shift in months. A new reasoning model, coding model, multimodal release, or agentic capability can change market perception almost overnight.
But the model layer no longer sits cleanly in the middle.
Model companies are moving down into tools and up into applications. OpenAI is not just a model API. Anthropic is not just Claude. Google DeepMind is not just a research lab.
These companies increasingly want to own the workflow around the model.
That is where the next layer comes in.
Layer 4: Data, developer tools, and model operations
This layer turns models into working systems.
It includes data platforms, orchestration, retrieval, evaluation, fine-tuning, observability, agent tooling, and developer workflows.
The traditional leaders include Databricks, Snowflake, Hugging Face, and Weights & Biases, with important players like LangChain, Pinecone, and MongoDB.
But this layer is under pressure.
Why? Because the model companies are building their own tooling.
Anthropic’s Claude Code is a clear example. It is not just a model feature. It is a developer workflow. It starts to absorb some of the value that might otherwise have gone to standalone tooling companies.
OpenAI is moving the same way with agents, tools, coding workflows, evaluations, and developer surfaces.
This does not mean independent data and tooling companies disappear. That would be too simplistic.
But it does mean the middle of the stack is getting squeezed.
The best analogy is what happened in cloud software. At first, there were many standalone tools around the cloud. Over time, the big platforms absorbed more of those functions natively. Some standalone companies still became huge, but they had to become much more specialized, more trusted, or more deeply integrated.
AI tooling may follow a similar pattern.
Layer 5: Apps, distribution, and devices
This is where AI reaches users.
The leaders include Microsoft, Google, Apple, and Meta, with important players like Amazon, Salesforce, Adobe, ServiceNow, and Oracle.
This layer is often underrated because it is less glamorous than frontier models. But distribution can matter as much as raw model quality.
Microsoft has Office, Teams, GitHub, Windows, Azure, and deep enterprise relationships.
Google has Search, Workspace, Android, Chrome, YouTube, and massive consumer reach.
Apple has devices, Apple silicon, on-device AI, privacy positioning, and Private Cloud Compute.
Meta has Facebook, Instagram, WhatsApp, Messenger, Quest, and Llama distribution.
Salesforce, Adobe, ServiceNow, and Oracle have deep enterprise workflow positions.
The lesson is blunt: the best model does not automatically win if another company owns the workflow where users spend their day.
Who is Most Vertical in AI?
Verticality matters because companies that span more of the AI stack can control cost, speed, supply, product integration, and distribution.
The question is not just who leads each layer, but who spans the most layers. Verticality creates strategic advantage through coordination and optimization.
Google: the broadest core-stack player
Google may be the closest thing to a full-stack AI company.
It has custom silicon through TPUs. It has Google Cloud. It has Gemini and DeepMind. It has Vertex and developer tooling. It has Search, Workspace, Android, Chrome, YouTube, and consumer distribution.
It is not a pure energy company, but across the core AI stack it is unusually complete.
This is Google’s structural advantage: it can connect research, infrastructure, product, and distribution in a way few others can.
The open question is not whether Google has the assets. It does.
The question is whether it can execute with enough urgency and product clarity to turn those assets into durable market leadership.
Amazon: more vertical than people think
Amazon is sometimes under-discussed in frontier AI conversations because it is not always perceived as having the leading model brand.
But Amazon is extremely important in the stack.
AWS is a central cloud platform. Trainium and Inferentia give Amazon custom silicon. Nova gives Amazon its own model family. Amazon also has massive distribution through commerce, enterprise cloud, devices, logistics, and media.
Amazon’s verticality is infrastructure-native.
It is not trying to look like OpenAI. It is trying to make AI a workload that runs through AWS and Amazon’s broader ecosystem.
That may prove very powerful.
Microsoft: distribution plus cloud
Microsoft’s strength is different.
It is less chip-deep than Google or Amazon, but it has extraordinary enterprise distribution.
Microsoft owns some of the most important knowledge-worker surfaces in the world: Office, Teams, Outlook, Windows, GitHub, LinkedIn, Dynamics, and Azure.
Its partnership with OpenAI also gave it early access to frontier AI capabilities and helped it move quickly into enterprise AI.
Microsoft’s verticality is not about owning every physical layer. It is about combining cloud, model access, workflow ownership, and enterprise trust.
That is a formidable position.
Meta: more vertical than the market often admits
Meta is not just a social app company and not just a model company.
It has massive consumer distribution through Facebook, Instagram, WhatsApp, Messenger, and Threads. It has Llama. It has enormous internal compute needs. It has custom silicon efforts through MTIA. It is also working with Broadcom on custom AI silicon.
Meta’s approach is different from Microsoft’s or Google’s. It has less enterprise workflow control, but more consumer distribution and a strong open-source model strategy.
The open-source dimension matters because it gives Meta influence beyond its own apps. Llama has become infrastructure for many developers and companies that do not want to be locked into closed model providers.
Apple: not a cloud power, but still strategically important
Apple is easy to underrate if you only look at public cloud and frontier model rankings.
But that misses Apple’s real leverage.
Apple has devices. It has Apple silicon. It has a tightly controlled operating system ecosystem. It has privacy as a brand asset. It has on-device AI. It has Private Cloud Compute.
Apple’s AI strategy is not about becoming AWS. It is about embedding AI into the personal computing layer.
That is a different kind of power.
The question for Apple is whether it can make AI feel indispensable on-device and across its ecosystem. If it can, it does not need to win the public model leaderboard to matter.
NVIDIA: the lower-stack king
NVIDIA may not own the consumer app layer, but it controls one of the deepest choke points in the system.
Its GPUs and software ecosystem are foundational to AI training and inference. CUDA, DGX, networking, systems, NeMo, NIM, and its broader AI infrastructure push make NVIDIA much more than a chip supplier.
NVIDIA is moving upward, but its real power is still below the model layer.
It is the company that many others need before they can compete.
That is leverage.
xAI / SpaceXAI: becoming a compute hyperscaler
This is one of the most interesting shifts.
xAI is not just trying to build Grok. Through Colossus and the broader SpaceXAI infrastructure story, it is moving into large-scale compute.
That matters because compute capacity is no longer just an input. It is becoming a product, a bargaining chip, and a strategic asset.
The Anthropic compute partnership is a signal. It suggests xAI / SpaceXAI can play not only as a frontier model participant, but also as a compute provider to other labs.
That is why I would place it in both the model layer and the compute-capacity layer.
Should SpaceXAI be counted in power and cooling? I would mark it as emerging or partial, not a core leader.
The idea of moving data centers to space is provocative and strategically interesting, but it is not yet the same as being a present-day leader in power and cooling for AI infrastructure. For now, I would treat that as an emerging thesis, not a current operating position.
Anthropic and OpenAI: model leaders moving into tools
Anthropic and OpenAI should both be shown as strong in frontier models and increasingly strong in developer tools.
Anthropic, in particular, deserves full credit in the tooling layer because Claude Code has become a serious developer workflow, not just a model interface. It is one of the clearest examples of a model company moving down into the tool layer.
OpenAI is also moving strongly in this direction through agents, coding tools, evaluations, fine-tuning, and developer surfaces.
Should either be shown as strong in compute cloud?
I would not show them as true compute hyperscalers.
Both rely heavily on infrastructure partners. They have access to enormous compute, and they shape demand for it, but they do not look like AWS, Azure, Google Cloud, Oracle, CoreWeave, or xAI / SpaceXAI in terms of owning and selling compute capacity as a core infrastructure business.
So yes: both OpenAI and Anthropic should be light or partial in compute, not solid.
That distinction matters.
Compute access is not the same as compute hyperscaler ownership.
Other hyperscalers not shown
The big AI hyperscaler conversation usually starts with:
- Microsoft Azure
- AWS
- Google Cloud
- Oracle OCI
- CoreWeave
- xAI / SpaceXAI
Depending on how broadly you define hyperscaler, you could also include:
- Crusoe, especially around AI data centers and energy-aware infrastructure
- Lambda, as a specialized GPU cloud provider
- Nebius, as an AI infrastructure player
- Voltage Park, as a compute provider
- IBM Cloud, though it is not leading the frontier AI cloud race in the same way
- Alibaba Cloud and Tencent Cloud, especially in China
- Huawei Cloud, especially in China’s AI infrastructure ecosystem
For a U.S.-centric or Western enterprise market map, Microsoft, AWS, Google, Oracle, CoreWeave, and xAI / SpaceXAI are the main names I would emphasize.
For a global AI infrastructure map, the China-based cloud and hardware players need their own section.
The Strategy of Choke Points
The AI race is increasingly a contest over choke points. The winners will not only build better models; they will control more of the system around the models.
The market is moving from novelty to infrastructure. In novelty markets, the best demo wins attention. In infrastructure markets, the deepest control points win economics.
In the early phase of generative AI, the market rewarded model breakthroughs. That made sense. GPT-4, Claude, Gemini, Llama, and Grok all changed the way people understood AI capability.
But as the market matures, advantage shifts from breakthrough alone to system control.
Who has power?
Who has chips?
Who has memory?
Who has compute?
Who has the best models?
Who has developer adoption?
Who has enterprise workflow access?
Who owns the device?
Who owns the customer relationship?
This is why the semiconductor layer deserves special attention.
The AI race is constrained less by model ideas than by access to the lower stack. NVIDIA still dominates accelerators. TSMC remains the manufacturing backbone. SK hynix, Micron, and Samsung matter because HBM has become a gating resource. ASML matters because advanced lithography is still a hard bottleneck. Broadcom matters in networking and custom accelerators.
The chip and memory layer deserves as much attention as the labs at the top.
In fact, sometimes it deserves more.
Because a model lab can invent a better architecture, but if it cannot get enough compute, memory bandwidth, and cluster scale, the idea may not become a product.
The new AI strategy question
The old question was:
Who has the best model?
The better question now is:
Who controls enough of the stack to shape economics, speed, access, distribution, and adoption?
That question gives you a more useful market map.
It explains why Google looks stronger when viewed across the full stack.
It explains why Amazon and Microsoft are both powerful, but in different ways.
It explains why NVIDIA’s position is so strategically important.
It explains why Apple belongs in the conversation even without a public cloud business.
It explains why xAI / SpaceXAI becoming a compute provider matters.
It explains why Anthropic and OpenAI are pushing deeper into developer workflows.
And it explains why the middle layer of data and model operations is becoming more contested.
The tool layer is where the next fight gets messy
One of the most interesting parts of the stack is the tooling layer.
This includes coding tools, evals, observability, data pipelines, vector databases, retrieval, fine-tuning, governance, agent frameworks, and model operations.
For the last couple of years, many startups and platforms have grown around the assumption that model companies would provide models, while everyone else would build the surrounding workflow.
That assumption is weakening.
Model builders want more of the workflow.
Anthropic does not want Claude to be only an API endpoint. Claude Code pushes Anthropic into developer workflow.
OpenAI does not want to be only a model provider. It is building agents, developer tools, and more of the application surface.
Google is doing the same through Gemini, Vertex, Workspace, and Android.
Microsoft is doing it through Copilot, GitHub, Azure AI, and enterprise workflow.
The tool layer is not going away. But it is being compressed.
Standalone tooling companies will need to prove they are not just temporary features waiting to be absorbed by the model platforms.
The survivors will likely be the ones with deep enterprise trust, strong data gravity, workflow specificity, governance strength, or cross-model neutrality.
In other words, “we are model-agnostic” may become more valuable for some customers, not less.
But it has to come with real depth.
A simple analogy: the AI stack is becoming the new oil industry
This analogy is imperfect, but useful.
In oil and gas, power does not sit only with the company that sells gasoline at the station.
It sits across the whole chain: exploration, drilling, refining, pipelines, shipping, storage, distribution, retail, and pricing.
AI is developing its own version of that chain.
Energy and cooling are the physical input.
Chips and memory are the extraction equipment.
Compute clouds are the refineries.
Frontier models are the high-value processed product.
Developer tools are the industrial machinery that lets companies use the product.
Applications and devices are the gas stations, cars, roads, and end-user experiences.
If you only look at the final product, you miss where the real leverage accumulates.
That is what is happening in AI.
The model is visible. The stack is powerful.
What executives should take away
For executives, board members, investors, and operators, the practical takeaway is this:
Do not evaluate AI companies only by model performance.
Ask better questions.
Where does this company sit in the stack?
Which bottleneck does it control?
Is it dependent on scarce chips, power, memory, or cloud capacity?
Does it own distribution?
Can it move across layers?
Is it building tooling around the model?
Does it have a path to enterprise adoption?
Can it lower cost over time?
Can it survive if model performance commoditizes?
These questions matter because the AI market is moving from novelty to infrastructure.
In novelty markets, the best demo wins attention.
In infrastructure markets, the deepest control points win economics.
The bottom line
The AI race is not just about who builds the smartest model.
It is about who controls the system around the model.
That means electricity. Cooling. HBM memory. Advanced chips. Networking. Cloud capacity. Frontier models. Developer workflows. Enterprise distribution. Consumer devices.
The companies with the most durable advantage will not necessarily be the flashiest.
They will be the ones that control the choke points others depend on.
That is why the AI competitive landscape is better understood as a stack war than a model race.
And that is why the next phase of AI strategy will be less about asking “who has the best model?” and more about asking:
Who owns enough of the stack to matter when the market gets hard?
