I am not a developer. Since May I have written 202,808 lines of code, shipped five health programs and the tools our team runs on, and turned hours of back-office work into minutes. That excites me, and it also worries me. Here is why, and what we are doing about it.
I wrote 202,808 lines of code this year. I cannot write one of them from memory.
Both of those sentences are true. It started in May, when I opened a file I did not understand to fix one small thing. I pushed the change and broke the site, and I ended up having to rebuild everything.
Four months later there are five health programs, the tools our team runs on, and thirteen code repositories. In September alone I wrote 79,208 lines.
To put that in perspective: before AI tools, a typical professional developer delivered about 10 to 50 lines of finished code a day, once planning, testing and fixing are counted, according to Steve McConnell’s Code Complete, the standard industry reference¹. Fred Brooks put the figure at about 10 a day in his classic The Mythical Man-Month². Over a 21-day working month, that is roughly 210 to 1,050 lines. In September I pushed 79,208, about 75 times the top of that range.
The two numbers are not measured the same way. The benchmark counts finished code across a whole project, while mine counts every line I committed, including code that was later rewritten, and lines of code have always been a rough way to measure work. But even with a generous discount, the gap is enormous, and it came from one person who would be, at best, an entry-level coder.
What the chart made possible
I did not learn to code. In OneVillage’s early days, in 2021 and 2022, I learned something worse: writing Jira stories. If you have never had the pleasure, a Jira story is a ticket that spells out, in painful detail, exactly what an engineer should build. “As a member, I want to reset my password, so that I can log in.” Then the acceptance criteria. Then the edge cases. Then the edge cases of the edge cases. I wrote more of them than I care to count and hated every single one.
It turns out they were the best training I never wanted. They taught me to break a problem into steps, to say exactly what “done” looks like, and to notice what is missing. Now I just tell the AI, in plain English, what I want the product to do. No zillion stories, no ticket numbers, no meetings where I nod along pretending to know what a story point is. The hardest part of the switch was letting go of the grudge. Today I describe exactly what a member should experience, look at what Claude builds, and check that it works the way it should. I do that ten or fifteen times a day. The clinical content inside every program is built by the OneVillage team and guided by our Medical Review Board, and we are grateful for the clinical expertise they bring to our development. That expertise is what makes the code worth anything.
Proven health programs rarely fail to reach people for scientific reasons. They fail because someone has to build the enrollment, scheduling, curriculum, outcome tracking, clinical notes and billing around them, and that has always needed a funded engineering team. AI took away most of that cost. These programs exist now because of it:
A ten-week cohort course
A learning portal for cancer survivors, with live group sessions, weekly lessons, homework, assessments and clinical notes in one place. Free for survivors in DC, Maryland and Virginia.
A chronic-condition course library
Twenty-five eight-week courses for members with chronic conditions, with weekly lessons, homework and a coach dashboard that shows each member's progress.
A self-paced skills course
Six one-on-one skill-building sessions for people with chronic conditions, with self-scheduling, lessons and progress tracking.
A heart-health course
An assessment-led learning path for first responders, with a participant portal and progress tracking.
A benefits decision tool
Employees answer a few questions and see what each of their employer's health plans would actually cost them, based on how they really use care. A navigator can then pick up the conversation from the same page.
The work around the work
The code is only the part you can count. The bigger change is in the work that used to fill my week. Contracts, onboarding, reporting and research each used to take hours of copying, checking and formatting, and they came out a little different every time. I have turned each one into a written, repeatable process that AI runs and a person approves. What took an afternoon now takes minutes, and it comes out the same way every time, whoever on the team runs it.
Contracting
Before: drafting each agreement by hand, then writing a cover email that explained the terms.
Now: the agreement and a plain-language summary of its terms are drafted together, ready for review and signature.
Onboarding
Before: building each new client's welcome packet, launch emails and guides one at a time, and the same for every new hire's paperwork.
Now: one process produces the full client launch kit or new-hire packet and records it in our client roster.
Reporting
Before: pulling usage numbers by hand into quarterly reviews and savings reports, then writing follow-up emails from memory.
Now: reports pull verified numbers, recheck every figure and keep a record of where each one came from. Follow-up emails are drafted from the call notes.
Research
Before: hours of reading to answer one market or competitor question.
Now: a sourced research brief in minutes, and a competitor tracker that updates itself every week.
A person still reads and approves everything before it reaches a client, a member or a new hire. What AI removed is the hours of assembly, not the care.
The other chart
Our engineering team keeps a dashboard of how much AI we use, measured in tokens, the units AI companies bill by. On 25 September our two engineers used roughly two billion tokens in a single day. A few months ago a normal day was a small fraction of that.
79,208: lines of code I wrote in September (Up from 3,008 in May)
~2 billion: tokens our engineers used on one day, 25 September (Our usage dashboard)
Both numbers measure the same thing: how much of OneVillage now rests on a small number of AI companies. Our products run on it, and so do our contracts, onboarding, reporting and research. If you took AI off my desk tomorrow, my chart would drop to zero the same week. That is the honest version of the story, and it is also a risk we have to plan for.
Why it looks like a bubble
None of this is a forecast. These are the warning signs in the public record, from central banks, consultancies, peer-reviewed research and the AI companies themselves, and we see several of them in our own work.
01. We pay less than it costs to make
OpenAI spent about $1.69 for every dollar it earned in 2025, according to financial documents reported by Fortune³. In the first quarter of 2026 it burned $3.7 billion against $5.7 billion in revenue⁴. Bain & Company estimates AI companies will need $2 trillion a year in revenue by 2030 to pay for the computing they are building, and expects the industry to fall about $800 billion short⁵. A discount that large is paid for by investors, and investors eventually want it back.
02. Quality does not hold steady
A Stanford and UC Berkeley study in the Harvard Data Science Review found that the "same" model can change a lot in a few months. GPT-4's accuracy on one math task fell from 84% to 51% between March and June 2023⁶. In September 2025 Anthropic published a review of three infrastructure bugs that had quietly made Claude's answers worse for several weeks⁷. We see the same pattern in our own work: a model is strong at launch and less reliable a few weeks later.
03. The lead is short
Stanford's 2025 AI Index found that the gap between the best paid model and the best free, open model shrank from about 8% to under 2% in a single year, and that the price of GPT-3.5-level performance fell more than 280-fold in two years⁸. DeepSeek reported in Nature that it trained its R1 reasoning model for about $294,000 on top of its base model⁹. In February 2026 Anthropic said three Chinese labs had pulled more than 16 million conversations out of Claude, through about 24,000 fake accounts, to train their own models¹⁰. When a rival can copy years of work that quickly, it is hard to see what holds up the valuation.
04. Markets are priced for perfection
In October 2025 the Bank of England warned that stock valuations were comparable to the peak of the dot-com bubble, with the five largest US companies making up close to 30% of the S&P 500, the highest share in 50 years¹¹. The IMF compared the AI boom to the late 1990s the same month¹², and in April 2026 it flagged "circular financing" among companies along the AI supply chain¹³. Meanwhile, MIT researchers found that 95% of companies they studied were getting no measurable return from their AI pilots¹⁴. If the returns do not catch up with the promises, the money can leave fast, and the dot-com crash did not stay inside tech.
The risk is not that the tools disappear. It is that the cheap part gets expensive and our customers get nervous in the same year.
What a pop would and would not mean
It would not take the technology away. Free, open models are now close behind the paid ones and keep getting cheaper to run⁸, so even in a bad year the tools stay within reach. They might be slower, and harder problems might take a minute instead of seconds, but the work still gets done. The IMF also notes that most AI spending so far has come from cash-rich companies rather than debt, which makes a 2008-style crisis less likely¹².
What would change is the price, the reliability, and the economy around us. Our buyers are employers. When markets fall, employers look for benefits to cut. That is the exposure that matters most to OneVillage, more than any model.
Roughly three outcomes are possible, and we should be ready for all of them.
Scenario 1: The skeptics are wrong
The technology keeps its promises, and the spending pays off. Tools keep getting better and cheaper.
Scenario 2: A hard pop
Valuations collapse, prices go up, and the slowdown reaches every industry, including our clients' budgets.
Scenario 3: A slow reset
Investors adjust their expectations. Growth slows but continues, and the discount on AI fades over time.
What could bring the cost down
Part of why AI is so expensive to run is where it runs. Data centers use enormous amounts of electricity and water. The International Energy Agency expects data centers worldwide to use about 945 terawatt-hours of electricity a year by 2030, more than double what they used in 2024 and more than all of Japan uses today¹⁵. In the US, Lawrence Berkeley National Laboratory estimated that data centers used 17 billion gallons of water for cooling in 2023, plus about 211 billion more through the power plants that supply them, and that those numbers could double or more by 2028¹⁶.
Some of the biggest names in technology think the answer is to move the computers off the planet. In the right orbit, a solar panel can be up to eight times more productive than on the ground and makes power almost around the clock, and heat can be released into space instead of cooled with water. Jeff Bezos predicted in October 2025 that gigawatt-scale data centers will be built in space within 10 to 20 years, and that they will eventually beat the cost of data centers on Earth¹⁷. Elon Musk is more aggressive. After SpaceX bought his AI company, xAI, in February 2026, he predicted that within two to three years space will be the cheapest place to run AI¹⁸, and SpaceX has filed plans for up to a million data center satellites¹⁹. Google has put numbers on the idea: if launch costs fall below $200 per kilogram, possibly by the mid-2030s, running computers in orbit could cost about the same as paying for power on the ground²⁰. The startup Starcloud has already trained an AI model on an Nvidia chip in orbit²¹.
Even at the same price, space wins: it spares the water and power our communities need, it isn’t stuck waiting on the grid, and it gets cheaper every time launch costs fall. Launch costs have already dropped by more than 95% since the Space Shuttle, and if the trend holds, computing in orbit keeps getting cheaper long after it matches the cost on Earth²².
Historical costs from NASA research; the mid-2030s projection is Google’s estimate, not a certainty.²²,²⁰
Not everyone agrees on the timing. Microsoft's president has said he would be surprised to see data centers move to orbit¹⁸, and Google's own math leaves out the cost of the chips and the buildings, with hard problems still unsolved, from radiation to moving data between satellites²⁰. But it shows that the cost of AI is not fixed. If power and water stop being the limit, AI could become cheap for real, not just subsidized by investors. That is the kind of cheap AI worth building toward, and one more reason to stay flexible about where our tools come from.
What industrial bubbles leave behind
Jeff Bezos, one of the biggest believers in space data centers, also calls today’s AI boom “a kind of industrial bubble.” A financial bubble, like the 2008 housing crash, mostly destroys wealth. An industrial bubble also wipes out a lot of investors’ money, but it leaves behind real infrastructure and inventions that everyone keeps using after the crash²³. History has several examples.
Railways, 1840s Britain
Investors piled into new railway companies in one of the biggest technology manias in history, and the crash that followed was one of the worst. But the projects approved in 1844 to 1846 still produced about 6,220 miles of track, which became the backbone of Britain’s rail network²⁴.
Fiber-optic cable, 1990s US
Telecom companies laid more than 80 million miles of fiber. Four years after the crash, 85 to 95% of it was still unused²⁵. Prices collapsed, and that cheap cable was later lit up to carry the streaming video and cloud services we use every day²⁶.
Biotech, 1990s
Bezos’s own example. Many biotech companies went bust, but the medicines developed during that boom are still saving lives today²³.
If AI follows the same pattern, many of the companies spending the most today may not survive. But the data centers, the chips, the cheaper open models, and perhaps the satellites will still be here. That is the bet we are making: use everything the boom builds, without depending on any single company that might not make it through.
How we hedge
We will keep using these tools hard. They are the reason a small team can serve people the market ignored. But we will use them as if the discount could end at any point.
1. Build things that last while building is cheap
Code we own, curricula, clinical documentation and outcomes data keep their value whatever happens to token prices. We should use this window to turn quick builds into lasting infrastructure, so our products live on one platform we control.
2. Don't depend on a single AI company
Our code lives in our own repositories, in standard languages any engineer can read. Any AI feature inside our products should be able to switch providers, including an open model we run ourselves. We should test that backup on a schedule, not just once it is needed.
3. Review harder as the tools change
If model quality can change from one week to the next, our review cannot. A person checks every change before it reaches a member, and we regularly stop adding features to tidy up and test what we have already built.
4. Treat AI spending like rent that can go up
We track compute spending every month, set a limit, and decide now what we would cut if prices went up five times. No product we sell should only make money because AI is temporarily cheap.
5. Sell results, not AI
Clients choose us for what our programs do for their people, not for the technology behind them. In a downturn, employers keep the benefits that save them money. Our coaches, navigators and clinicians, and the costs they save, are the product. AI is how we deliver it more cheaply.
6. Grow on revenue, not on the AI story
Our growth is all inbound and paid for by customers. We should keep it that way, keep more cash on hand than feels necessary, and raise money on revenue we have earned, not on a valuation that depends on AI excitement.
The short version. Use the tools as much as we can while they are cheap, and put what they produce into things we own. Keep a backup ready. Keep a person reviewing every change. Sell the outcome, not the technology.
Why both lessons matter
The chart teaches two lessons. The first is that anyone who deeply understands a group of people the market has ignored can now build for them. The second is to build as if the discount could end tomorrow. Both are true, and they depend on each other.
Here is what I mean. The first lesson is the opportunity: cheap AI is what let us bring a cancer program to survivors who could never have traveled for it. The second lesson is the discipline: assume that cheap AI will not last. The discipline is what protects the opportunity. If we build products that only work while AI is cheap, then the day prices go up or a model gets worse, those programs stop, and the survivors, first responders and employees who depend on them lose them. If we instead use this window to build things we own, can run on more than one AI provider, and can afford at higher prices, the programs keep running whatever happens to the AI market. Building as if the discount could end is how we make sure the people we serve do not lose access when it does.
I started out trying to fix one small thing and ended up rebuilding everything. Start anyway. Just build things that will still be standing if the price changes.
Sources
- S. McConnell, Code Complete, 2nd ed., Microsoft Press, 2004 (industry average 10 to 50 lines of delivered code per person per day); summarized by Coding Horror.
- F. P. Brooks, The Mythical Man-Month, Addison-Wesley, 1975.
- Fortune, reporting on OpenAI financial documents, 12 Nov 2025.
- The Information, 16 Jun 2026, as reported by PYMNTS, “OpenAI Ran Through $3.7 Billion in Q1 2026.”
- Bain & Company, Global Technology Report 2025, Sep 2025.
- Chen, Zaharia & Zou, “How Is ChatGPT’s Behavior Changing Over Time?” Harvard Data Science Review 6(2), 2024.
- Anthropic, “A postmortem of three recent issues,” Sep 2025.
- Stanford HAI, The 2025 AI Index Report, Apr 2025.
- DeepSeek-AI, “DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning,” Nature, 17 Sep 2025; cost figure as reported by CNN.
- TechCrunch, “Anthropic accuses Chinese AI labs of mining Claude as US debates AI chip exports,” 23 Feb 2026.
- Bank of England, Record of the Financial Policy Committee meeting, 2 Oct 2025.
- International Monetary Fund, Global Financial Stability Report, Oct 2025, Chapter 1; see also CNBC, 9 Oct 2025.
- International Monetary Fund, Global Financial Stability Report, Apr 2026.
- MIT NANDA, The GenAI Divide: State of AI in Business 2025, Jul 2025.
- International Energy Agency, Energy and AI, Apr 2025.
- Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report, Dec 2024.
- The Japan Times, “Data centers in space? Jeff Bezos says it’s possible,” 4 Oct 2025.
- TechRadar, “Musk insists that ‘the lowest cost way to generate AI compute will be in space’ within three years,” Feb 2026.
- Data Center Dynamics, “SpaceX files for million satellite orbital AI data center megaconstellation,” 2026.
- Google Research, “Exploring a space-based, scalable AI infrastructure system design,” 4 Nov 2025.
- CNBC, “Nvidia-backed Starcloud trains first AI model in space,” 10 Dec 2025.
- H. W. Jones (NASA Ames), “The Recent Large Reduction in Space Launch Cost,” 48th International Conference on Environmental Systems, 2018.
- CNBC, “Jeff Bezos says AI is in an industrial bubble but society will get ‘gigantic’ benefits from the tech,” 3 Oct 2025.
- A. Odlyzko (University of Minnesota), “Collective Hallucinations and Inefficient Markets: The British Railway Mania of the 1840s,” 2010.
- Fortune, “Here’s what went down 25 years ago that ultimately burst the dot-com boom,” 28 Sep 2025.
- Communications of the ACM, “Dark Fiber Is Lighting Up.”
Lines of code are non-merge commits authored under my GitHub account across OneVillage repositories, 27 May to 30 September 2026, measured with the GitHub API. Lockfiles and other machine-generated dependency files are excluded; code counts .ts, .tsx, .js, .py, .sql, .css, .html and .sh files. Token usage comes from our engineering team's usage dashboard, as of 1 October 2026; the daily figure is approximate. Nothing here is investment advice.
