This is Part 2 in a three-part series on AI's real competitive barriers: where they're being built, who benefits, and what founders and business leaders should do about it.
In Part 1: The Compute Chokehold & Who Owns the Corpus, I explored how compute and data, not model quality, are already the decisive competitive barriers in AI.
Distribution Is Destiny
Let’s say you’re a startup. You’ve got a great model. Maybe it’s even better than GPT-5 on some tasks. Congrats. Now what?
Microsoft has OpenAI baked into Windows, Office, GitHub, Azure, Bing, and Copilot. That’s not a product integration; it’s an occupation. Hundreds of millions of users hitting OpenAI endpoints every day without even knowing it. The defaults are set. The habits are forming.
Amazon is playing both sides here too. Bedrock hosts multiple foundation models across AWS, but the new OpenAI deal makes AWS the exclusive third-party cloud distributor for OpenAI’s enterprise platform. Google’s pushing Gemini into Search, Gmail, Workspace, Android, and YouTube. Every surface they own becomes a distribution channel.
Same playbook as always. Microsoft did it with Internet Explorer. Google did it with Chrome. The AI version is just faster and more aggressive. Founders with impressive models can’t get distribution and are hustling for customers one developer at a time while the incumbents are shipping to every enterprise and consumer on the planet simultaneously. That’s not a fair fight.
I recently spoke with a senior exec at a major investment bank betting big on AI. They openly expressed a strong bias towards enterprise-class vendors like OpenAI, and warned that their onboarding and vendor qualification process could take 18 months. That timeline is probably incompatible with the bandwidth, or runway, of a typical startup. But here’s the kicker: even if a startup had an offering they needed and it couldn’t be sourced elsewhere, their first move would be to attempt to recreate the capability in-house. And failing that, they’d push OpenAI to prioritize that feature. The startup isn’t even third in line. It’s a last resort.
The same pattern is emerging in healthcare, just with different acronyms. OpenAI has rolled out ChatGPT Health and an enterprise “OpenAI for Healthcare” stack, with HIPAA‑aligned APIs, Business Associate Agreements and integrations designed to sit inside hospital systems, payers, and health‑tech vendors. If they succeed, patients, clinicians, and third‑party tools will all interact through the same substrate, turning OpenAI from “a model vendor” into the default orchestration layer for how health data is queried, summarized, and acted on in practice. In that world, “distribution” doesn’t mean winning another EHR pilot; it means being the fabric the EHR itself calls out to when anything intelligent needs to happen.
Anthropic is making a parallel bet, but at a different layer. Rather than positioning Claude as the orchestration layer above workflows, they’re wiring it directly into data sources: CMS coverage databases, ICD-10 and NPI registries, ClinicalTrials.gov, PubMed, Medidata, and other clinical data platforms through a growing connector ecosystem. Their Claude for Healthcare suite, launched in January 2026, goes further: it includes HIPAA-ready APIs and built-in agent skills for drafting clinical trial protocols to FDA requirements and preparing regulatory submissions.
The strategic logic is that the real advantage lives in the pipes. If Claude becomes the default interface for querying claims, trials, and literature, the reasoning increasingly happens through Claude rather than against raw, unintegrated sources. Where OpenAI is building distribution through product surface area, Anthropic is building it through plumbing, but the end state is the same: structural dependency.
Meanwhile, pharma companies are also partnering directly with AI infrastructure giants. In January 2026, Eli Lilly and Nvidia committed up to $1 billion over five years to build a joint AI lab in the San Francisco Bay Area. This will marry Lilly’s proprietary drug data with Nvidia’s computing muscle to create a platform of biology-specific foundation models for drug discovery. The pitch is “closed-loop discovery”: computational predictions feeding directly into lab experiments and back again, compressing timelines that used to take years. If it works, this becomes a structural advantage few outsiders can replicate, relegating everyone else to bit roles. In March, Roche announced the largest hybrid-cloud AI factory in pharma, with over 3,500 NVIDIA GPUs including 2,176 new Blackwell units, to accelerate drug and diagnostic development. Two of the world’s largest pharma companies, weeks apart, both locking in NVIDIA partnerships.
Regulation as Warfare
At first, I found it perplexing that the loudest voices calling for AI regulation were the leading AI companies themselves. Then I realized this pattern appears across many industries. Incumbents tilt the playing field to their advantage by influencing and shaping the regulatory environment. There’s a term for this: regulatory capture. And it’s playing out in real time.
When Anthropic or OpenAI or Google testify about catastrophic risks and advocate for licensing regimes, training thresholds, and mandatory safety evaluations, they’re not wrong about the risks. But every requirement they propose is a requirement they can meet and their competitors probably can’t.
The EU AI Act will hit startups hardest. The compute thresholds alone will define “frontier” in a way that locks in the current leaders. And even where SME accommodations exist, navigating the regulatory maze takes resources that startups don’t have and incumbents do.
Of course, safety matters. But we should be clear-eyed about who benefits when “safety” gets operationalized as “regulations that only five companies can afford to follow.” The moat isn’t just compute and data; it’s also policy. And the incumbents are building that barrier in plain sight.
The Open-Source Mirage
“But what about Llama? What about Mistral? What about DeepSeek? Open-source is catching up!” It’s not entirely wrong. Meta’s releasing good models. Mistral is punching above its weight. And DeepSeek genuinely shook the industry: US tech stocks tumbled when R1 dropped in early 2025, with Marc Andreessen calling it a ‘Sputnik moment.’ Their latest model, V3.2, matches GPT-5 on reasoning benchmarks, costs a tenth of the price to run, and they open-sourced the whole thing under an MIT license. In January, Moonshot AI released Kimi K2.5, a trillion-parameter open-source model that outperforms GPT-5.2 on several benchmarks.
So yes, open-source is real. But performance parity is not competitive parity.
DeepSeek is backed by High Flyer, a well-capitalized Chinese hedge fund, making it hardly a scrappy garage project. And despite the technical achievement, enterprise adoption remains constrained by security concerns, data sovereignty questions, and China-origin risk. Case in point: Texas banned state employees from using it. That’s not a frictionless path to market dominance.
The real barrier is capital, infrastructure, and institutional trust, not whether your model is open. Benchmark parity alone doesn’t get you through the door. Try to sign an enterprise contract with a Fortune 500 company and they’re going to ask about your SLAs, your compliance certifications, your data residency, and your liability coverage. They’re going to ask if you’ll still exist in two years. Open-source models don’t have exclusive training data, global-scale inference infrastructure, enterprise sales teams, regulatory lobbying operations, or default distribution in products with a billion users. Llama is good. But it’s an anomaly, not a template for the ecosystem. Meta can afford to give it away because Meta makes money elsewhere.
The same dynamic is visible in the life sciences. Arc Institute, a nonprofit funded by Stripe co-founder Patrick Collison, trained Evo 2, a 40-billion-parameter genomic foundation model, on over 2,000 NVIDIA H100 GPUs and released it openly. The result is genuinely frontier work, published in Nature. But it took philanthropic backing from a billionaire funder and a direct NVIDIA partnership to get there. Open-source is not an alternative to the infrastructure economics of AI. It is subject to the same constraints: scarce compute, expensive training runs, and access to capital.
The sentiment in the open-source community is optimistic, but investors aren’t buying it. Money is flowing to the frontier labs, not the open alternatives. In 2025, frontier AI labs raised over $80 billion, with most going to five companies. That looks like a bet on an oligopoly, not a competitive market.
Coming up in Part 3:
In Part 3 of this series, I will discuss what the endgame actually looks like, and what founders and enterprise leaders should do about it. The Great Bifurcation, and practical advice for the boardroom.
If you found this article valuable, share it with a founder, investor or executive navigating the AI landscape.
Thanks for reading Backed By Reality.
Subscribe for free to receive new posts.
Your point on distribution being the real moat is exactly right, and we’re already seeing it play out in market data.
Even outside the core model providers, companies with strong existing distribution are outperforming expectations. Canva is a good example, consistently ranking alongside or even ahead of some established LLM players in usage and adoption despite not being a frontier model lab.
That aligns with what Andreessen Horowitz has been highlighting: the winners aren’t necessarily the ones with the best models, but the ones embedding AI into products people already use daily.
In that sense, AI isn’t just a model race, it’s a distribution land grab.