This is Part 5 of the Vertical Wall series. Read Parts 1–4 here.
Someone is building bridges.
We've spent four parts of this series staring at the wall. The talent divide that's splitting the workforce into AI-haves and AI-have-nots. The enterprise gap where Fortune 500s pour billions into tools their employees can't use. The geographic concentration funneling all the money and all the talent into fifteen zip codes. The institutional paralysis universities, governments, workforce programs frozen while the ground moves underneath them.
The wall is real. Everything we've documented is happening right now, in real time, at a pace that makes previous technology transitions look leisurely.
But here's the thing about walls: people don't just stare at them.
Free School
In March 2026, Anthropic launched the Anthropic Academy. Thirteen free courses covering the full AI stack from prompting fundamentals to building production applications. No paywall. No waitlist. No prerequisites beyond curiosity and a browser.
This from the company that positions itself as the cautious one. The safety-first lab. The people who publish papers about AI risk and then, apparently, also build the best free AI curriculum on the internet.
Contradictory? Maybe. But the courses are real. And they're good.
Anthropic isn't alone. OpenAI and Anthropic both launched college-focused education programs in early 2025. Google has been flooding Coursera with free AI courses and now offers an AI Essentials certification no degree required, no tuition, just you and your browser. Microsoft launched an AI Skills Initiative through LinkedIn Learning, leveraging the professional network where 900 million people already have profiles.
Read that list again. Anthropic. OpenAI. Google. Microsoft. The four most powerful AI companies on the planet are all racing to give away education for free.
Why?
The cynical read: they need users. More people who understand AI means more people buying API calls, cloud compute, enterprise licenses. Education is customer acquisition dressed in altruism.
The generous read: they see the wall too. And they know that AI adopted by 15% of the workforce isn't a revolution it's a luxury. These companies need broad-based adoption to justify their valuations, their hiring plans, their existence. If AI becomes a thing that only knowledge workers in expensive cities use, the entire narrative collapses.
The honest read is probably both. Motives don't have to be pure to produce real results. And the results are real: thousands of hours of free, high-quality AI education that didn't exist eighteen months ago.
Andrew Ng said it years ago: "AI is the new electricity." He was talking about potential. The free education wave is the part where someone starts wiring the houses.
Open Source as Infrastructure
Education gets you knowledge. Open source gets you tools.
Meta's Llama has become the Linux of AI. Over 85,000 derivative models on Hugging Face. Llama 4 Scout and Maverick are natively multimodal and free to download. Mark Zuckerberg has been saying the quiet part loud: "Open-source is how we make sure AI isn't just something that benefits a few big companies."
He's not wrong. Whatever you think about Meta's motives and there's plenty to think about the effect is undeniable. An Indian startup can now build a Hindi healthcare chatbot on Llama without paying a dollar in licensing fees. A team in Lagos can deploy an agricultural advisor in Yoruba using the same foundational model that powers applications at Fortune 500 companies.
And Llama isn't the only game. DeepSeek, Mistral, and a growing constellation of open-weight models have created a genuinely competitive ecosystem. Chinese open-source model downloads surpassed Western ones in August 2025 a sentence that should reshape how you think about AI geopolitics.
But here's what really changed the game: the tools that let you run these models without a data center.
Ollama. LM Studio. llamafile. These are applications that let you download and run AI models on a laptop. No API key. No cloud bill. No corporate IT department granting access. You download the app, pull a model, and start using it. The entire interaction happens on your machine.
This is a bigger deal than it sounds. We talked in Part 2 about the enterprise gap companies buying AI tools their employees can't use because of security reviews, procurement cycles, and IT policies that move at institutional speed. Ollama bypasses all of that. Not officially, not with permission, but it bypasses it. A curious employee can have a local AI running on their work laptop in ten minutes.
McKinsey found that employees are three times more likely than their leaders expect to already be using AI daily. This is why. The tools didn't wait for permission.
Hugging Face now hosts over 500,000 models. Half a million. Community-built, community-documented, community-improved. Yann LeCun, Meta's chief AI scientist, has been saying it for years: "The history of technology shows that open platforms always win in the long run."
The long run might be shorter than anyone expected.
Consider what this ecosystem means in practice. A developer in Nairobi and a developer in New York now have access to the same foundational models, the same fine-tuning frameworks, the same deployment tools. The New York developer has faster internet and better hardware. The Nairobi developer has lower costs and local context that no San Francisco lab can replicate. For the first time in the history of technology, the starting line isn't determined entirely by geography and capital. It's still uneven we'll get to that but the gap between "possible" and "accessible" has never been smaller.
From Lagos to Louisville
The most interesting bridges aren't being built in San Francisco.
AI Saturdays Africa runs free weekly workshops in Lagos. The format is simple: show up on Saturday, learn AI, build something. Participants go from zero to deploying working models in months. No tuition. No admissions committee. No venture capital required just a room, a Wi-Fi connection, and people who want to learn.
In India, startups are building Hindi healthcare chatbots and Tamil agricultural advisors on top of Llama. These aren't research projects or demo apps. They're tools that help farmers check crop prices and patients navigate symptoms in languages that the major AI companies barely support. The open-source models gave them a foundation. The local knowledge made it useful.
Dario Amodei, Anthropic's CEO, wrote in his essay "Machines of Loving Grace" that "AI could dramatically accelerate economic development in the developing world... but this requires deliberate effort." Lagos and Chennai aren't waiting for deliberate effort from Silicon Valley. They're making their own.
But you don't have to fly to Lagos to find bottom-up adoption.
There's a teacher in rural Kentucky who found ChatGPT through TikTok. Not through a district training. Not through a professional development workshop. Through her For You page. Now she creates personalized lesson plans, generates differentiated assessments, and saves hours every week. No one told her to do this. No institution enabled it. She just found a tool and started using it.
She's not alone. Microsoft's research shows 44% of teachers now use AI for research, 38% for lesson planning, 37% for creating tests. Most of them figured it out on their own. The institutions they work for are still debating whether to allow it.
This is the pattern, repeated everywhere: individuals moving faster than the organizations that employ them. Workers crossing bridges that their companies haven't approved yet. The McKinsey data on shadow AI adoption isn't a warning about compliance risk it's evidence that the bridges work.
People are voting with their browsers.
And it's not just teachers. The pattern shows up in every industry we've examined in this series. Nurses using AI to decode insurance prior authorizations. Electricians checking code compliance on their phones. Real estate agents generating listing descriptions and market analyses in minutes instead of hours. None of them were trained. All of them are more productive. Most of their employers have no idea.
What's Still Missing
Before we get too optimistic, let's be honest about what free courses and open-source models don't solve.
Compute. You can run a 7-billion-parameter model on a laptop. You cannot train one. You cannot fine-tune a large model on custom data without serious GPU access. The gap between "using AI" and "building with AI" still has a hardware gate, and that gate has a price tag. NVIDIA's dominance in AI chips means the cost of serious compute hasn't democratized the way the models themselves have.
Data infrastructure. A free course teaches you to prompt. It doesn't teach your organization to build a data pipeline, clean its datasets, or create the feedback loops that make AI actually useful in context. The hardest part of enterprise AI adoption isn't the model it's everything around the model.
Implementation support. The Kentucky teacher figured it out. But she's a self-starter who happened to find the right TikTok. What about the millions of workers who need structured support, hands-on guidance, and institutional backing to cross the same bridge? Free education is necessary but not sufficient. Someone still has to help people walk across.
The permission gap. We covered this in Part 2: organizations are simultaneously buying AI tools and blocking their employees from using them. The bridges exist, but many institutions have posted "No Trespassing" signs at both ends. Until leadership catches up to their own workers, the individual bridges will remain guerrilla crossings rather than sanctioned pathways.
PwC's data tells a nuanced story: AI skills command 56% wage premiums right now, but the pool is expanding rapidly through free courses. That premium will compress. The question is whether it compresses because everyone gains the skills a good outcome or because the market decides the skills only matter at the top, leaving the rest behind. Again.
The Take
We started this series with a wall. A vertical divide between people who use AI and people who don't, between companies that have figured it out and companies pretending they have, between cities that attract AI talent and everywhere else.
Five parts later, here's what I believe:
The wall is real but not inevitable.
The tools to tear it down exist. They're free. They're open-source. They run on hardware people already own. The education to use them is available from the companies that build the most powerful AI systems on Earth, and it costs nothing.
Millions of people are already using them without institutional permission, without formal training, without anyone's strategic initiative making it happen. The Kentucky teacher didn't wait for her district's AI policy. The developers in Lagos didn't wait for Silicon Valley to notice them. The employees in every McKinsey survey didn't wait for their CTO's roadmap.
They just started.
That's the hopeful part. The grassroots adoption is real, it's accelerating, and no policy document or institutional delay can fully stop it.
But here's the honest part: grassroots isn't enough.
Individual bridges help individuals. They don't close the systemic gap. They don't retrain the 40-year-old warehouse manager whose job is changing. They don't equip the community college in rural Mississippi. They don't solve the compute access problem or the data infrastructure problem or the problem of leaders who think buying a Copilot license counts as a strategy.
The bridges exist. The question is whether institutions companies, universities, governments, the entire apparatus of organized society will walk across them. Or whether they'll keep staring at the wall, commissioning reports about the wall, forming committees to study the wall, while the people who figured it out on their own pull further and further ahead.
We've seen this movie before. The internet created a divide. Mobile created a divide. Cloud computing created a divide. Each time, the tools eventually became cheap enough and simple enough that most people caught up. But "eventually" took a decade or more, and the people who were late paid for it in wages, opportunities, and relevance.
AI is moving faster than any of those transitions. The bridges are being built faster. But the wall is going up faster too.
Every month that passes without institutional action, the early adopters compound their advantage. The 56% wage premium doesn't sit still it reshapes hiring, promotions, entire career trajectories. The companies that figured out AI deployment don't just outperform their competitors they attract the talent that makes the gap permanent. The cities that built the ecosystem don't just grow they become the only places where certain careers are possible.
The bridges are free. Walking across them is not costless it takes time, effort, and the willingness to feel stupid for a while. But the toll is measured in weeks, not years. In curiosity, not credentials.
Which one wins is not a technology question. It's a choices question.
A product manager in San Francisco opens her laptop at 7:30 AM. She's still using AI for everything. She's faster than ever.
Eight hundred miles east, the HVAC technician in Louisville has started his morning differently. His kid showed him a YouTube video about a free app that runs on his phone. He asked it a question about a compressor diagnostic code he couldn't figure out. It gave him the answer in thirty seconds.
He didn't tell his boss. Nobody trained him. No institution made it happen.
He just walked across a bridge.
The wall is still there. But the first cracks are showing. And on the other side, people are already building.
This concludes the Vertical Wall series. Parts 1–4 covered the talent divide, the enterprise gap, the geography problem, and the institutional freeze. If you've followed along: thank you. Now go build a bridge.