AI margin squeeze: 8 companies ranked by economic resilience

Token prices are in freefall. Compute costs are not.

That mismatch is the central tension Steve Hou, head of research at Silicon Data, laid out in a recent Bloomberg Merryn Talks Money episode. The basic problem: AI companies sell tokens (queries, outputs, API calls) at prices that are collapsing due to competition, while the cost of the GPUs and data centers needed to produce those tokens keeps rising. The result is a margin squeeze that will separate structurally sound AI businesses from ones running on narrative and funding rounds.

For anyone thinking about AI brand visibility and GEO, this matters more than it might seem. If the economics of AI inference collapse, the mix of models powering AI search changes fast. The AI engine your content is optimized for today might be a smaller player next year, or might have pivoted its product to survive.

So I ranked 8 major AI companies by their ability to weather this squeeze. Not by hype. By structure.

Ranking methodology

Four criteria, each weighted roughly equally:

1. Revenue diversification (25%): Does the company have meaningful income streams beyond raw API token sales? Enterprise contracts, subscriptions, hardware, and adjacent software all count.

2. Compute ownership or lock-in advantage (25%): Does the company control its own silicon, have preferential supply agreements, or run proprietary hardware that reduces cost per token structurally?

3. Funding runway and profitability trajectory (25%): How much cash does the company hold, and is the trajectory toward profitability credible given current burn rates?

4. Demand moat (25%): Is there evidence of genuine user lock-in, unique data flywheel effects, or enterprise stickiness that makes switching painful enough to protect pricing power?

By the numbers

Token prices from major LLM providers have dropped by an estimated 90% in 18 months as of mid-2026, driven by intense competition from open-source models and new entrants (The Economist, 2025). That means every dollar of API revenue a company earned in early 2024 now buys roughly ten cents of equivalent compute time at the same price point. Margins that looked acceptable then have evaporated.

Microsoft committed $80 billion to AI data center construction in fiscal year 2025 alone (Microsoft Blog, 2025). That scale of capital expenditure only makes sense if token demand grows by orders of magnitude, which is the bet the entire industry is making in unison.

OpenAI reportedly reached $10 billion in annualized revenue in mid-2025 but was simultaneously burning through cash at a rate that made profitability a 2026-or-later story (The Information, 2025). Revenue and profit are very different things when your infrastructure bill scales with your revenue.

An estimated 70% of enterprise AI workloads will run on some form of fine-tuned or self-hosted open-source model by 2027 according to Gartner projections (Gartner, 2025). That is a direct threat to the API-revenue model that most closed-model companies depend on.

Google's TPU infrastructure gives it a structural cost advantage estimated at 30-40% lower inference cost per token versus equivalent GPU-based competitors, based on public benchmarks from MLCommons (MLCommons, 2024). That is not a gap competitors can close quickly.

The 8 companies ranked

1. Google DeepMind / Google Cloud

Google's vertical integration is the clearest structural advantage in this market. TPU v5 and v6 hardware gives Google inference costs no GPU-dependent competitor can match. Revenue diversification is extreme: AI revenue flows through Cloud, Search, Workspace, and Android simultaneously. The weakness is organizational: Google has historically struggled to commercialize AI research at the speed its research output would suggest is possible. But on margin resilience? Top of the list.

Specific strength: Proprietary silicon + massive existing distribution = the lowest cost-to-serve ratio of any major player.

Specific weakness: Slow internal product cycles mean competitors capture enterprise mindshare even when Google's underlying model quality is competitive.

2. Microsoft (OpenAI partnership + Azure)

Microsoft does not build the frontier model itself, but it has the distribution that makes margin squeezes less painful. Azure AI revenue is bundled into enterprise contracts that are sticky and multi-year. Copilot integration across Office 365 means AI usage generates subscription revenue, not just token revenue. The risk is dependency: if OpenAI's unit economics collapse entirely, Microsoft's AI narrative takes a real hit even if its financials hold.

Specific strength: Enterprise contract stickiness insulates it from spot token price competition.

Specific weakness: Heavily dependent on OpenAI's model quality remaining competitive, which is not guaranteed.

3. NVIDIA

NVIDIA is not an AI model company, but it is the most important company in this analysis because it captures margin from both sides of the squeeze. Whether token prices rise or fall, someone is buying H100s and B200s to build the infrastructure. NVIDIA's CUDA moat means switching to alternative silicon requires years of software re-engineering. The risk is a real one: if hyperscalers succeed in custom silicon (Google already has, Amazon is making progress with Trainium), NVIDIA's dominance softens over a 5-7 year horizon.

Specific strength: Earns margin regardless of which AI company wins the token war.

Specific weakness: Custom silicon from hyperscalers is a slow but real structural threat.

4. Amazon (AWS Bedrock + Trainium)

AWS Bedrock offers a model-agnostic AI platform that is less exposed to any single model's pricing or quality issues. Trainium 2 chips give Amazon a path to lower inference costs without NVIDIA dependency. Enterprise relationships are the deepest in cloud. The weakness: Amazon has not yet produced a frontier model that competes on capability benchmarks, which limits its ability to capture the high-margin enterprise AI consulting and fine-tuning work.

Specific strength: Model-agnostic positioning means it benefits from commoditization rather than suffering from it.

Specific weakness: No competitive frontier model capability of its own.

5. Anthropic

Anthropic is the most interesting mid-tier case. Claude's reputation for safety and reliability has created genuine enterprise stickiness, particularly in regulated industries like legal, finance, and healthcare. The company has raised enormous sums and secured significant AWS investment. But it is still burning cash at scale with no hardware advantage and no distribution moat comparable to the hyperscalers. The margin squeeze hits Anthropic harder than companies above it on this list. Its path to survival is acquisition or becoming the preferred safety-tier model for regulated enterprise, which is a real but narrow lane.

Specific strength: Safety reputation creates pricing power in regulated verticals that pure-performance models cannot easily access.

Specific weakness: No hardware advantage, high burn rate, dependent on continued investment rounds.

6. OpenAI

OpenAI has brand recognition that no other pure-play AI lab can match, and ChatGPT's consumer install base is a genuine data flywheel. But the economics are difficult. Token prices are dropping, compute costs are not falling fast enough, and the company's ambition to build its own hardware is years from maturity. The for-profit restructuring signals that investor pressure for returns is real. OpenAI's best margin protection is its consumer brand creating demand for premium-priced enterprise tiers, but that margin is compressing.

Specific strength: Consumer brand and ChatGPT distribution create enterprise pipeline that pure API companies cannot match.

Specific weakness: Spending at a scale that makes profitability genuinely uncertain without hardware cost reduction.

7. Mistral AI

Mistral has built a clever positioning as the open-weight model company that enterprises can self-host, removing API cost entirely. That is a smart response to the margin squeeze: let customers absorb the compute cost and charge for enterprise support, fine-tuning, and compliance tooling instead. The weakness is scale: Mistral does not have the resources to push capability frontiers, so it will always be one or two generations behind the leading closed models on raw benchmark performance.

Specific strength: Open-weight model positioning turns the margin squeeze into a sales pitch.

Specific weakness: Cannot compete on frontier capability without frontier compute budgets.

8. Cohere

Cohere targets enterprise specifically, which is where the margin logic is most defensible. Its models are not consumer-facing, which means it avoids the brutal consumer API pricing competition. Enterprise fine-tuning and retrieval-augmented generation (RAG) deployments create sticky relationships. The problem is that every hyperscaler is now pitching the same enterprise story with more distribution, more existing relationships, and more compute. Cohere's window to establish durable enterprise relationships is narrow.

Specific strength: Enterprise-only focus means it is never competing in the consumer token price war.

Specific weakness: Hyperscalers are encroaching on exactly the enterprise RAG market Cohere needs to own.

Comparative scorecard

Scoring is based on publicly available financial reporting, infrastructure announcements, and analyst estimates. Each criterion is scored 0-100% for quantitative measures and 1-5 stars for qualitative assessments.

Company	Revenue diversification	Compute advantage	Funding runway	Demand moat	Overall resilience
Google DeepMind	95%	★★★★★	95%	★★★★★	★★★★★
Microsoft	90%	★★★★☆	90%	★★★★★	★★★★★
NVIDIA	85%	★★★★★	90%	★★★★☆	★★★★★
Amazon AWS	88%	★★★★☆	92%	★★★★☆	★★★★☆
Anthropic	40%	★★☆☆☆	60%	★★★☆☆	★★★☆☆
OpenAI	55%	★★☆☆☆	55%	★★★★☆	★★★☆☆
Mistral AI	35%	★★★☆☆	45%	★★☆☆☆	★★☆☆☆
Cohere	50%	★★☆☆☆	50%	★★★☆☆	★★☆☆☆

What this means for GEO practitioners

The margin squeeze is not just a finance story. It is a platform stability story.

If Anthropic's economics deteriorate, Claude gets less investment in search-facing features. If OpenAI pivots hard toward consumer subscriptions to protect margin, its API quality and reliability for enterprise search integrations could shift. The AI engines powering the queries you're optimizing for are not static platforms. They are companies fighting for survival in a brutal cost environment.

This is exactly why source authority matters more than platform-specific tricks. If you're building GEO visibility by gaming one specific engine's quirks, you're exposed to that engine's business model risk. If you're building authoritative content that earns citations across multiple AI engines, your GEO equity survives platform turbulence.

Tools like winek.ai track visibility across ChatGPT, Perplexity, Gemini, Claude, Grok, and DeepSeek simultaneously for exactly this reason. Single-engine concentration is a business risk right now, not just a measurement limitation.

The companies with the deepest compute advantages and the most diversified revenue streams will shape what AI search looks like in 2027 and 2028. Google and Microsoft are structurally dominant. Everyone else is playing a more uncertain game.

The brands whose GEO strategies account for that uncertainty will be better positioned regardless of which labs survive the squeeze.