G.L.A INVE$T RE$EARCH w COACH KYLE THE LARGEST INNOVATION BOOM / ALL ABOUT COST DECLINES Sunday, September 20, 2026
Research. Conviction. Exponential Growth.
Wall Street keeps asking if this is a bubble.
The Data says otherwise.
As we know, we are in the early stages of the largest innovation boom since the early 1900’s.When the cost of intelligence, electrons, physical work, trust, and biology all collapse at the same time, demand does not rise in a straight line. It explodes.That is what our research is saying. That is the game board, and that is why we are investing in the best of the best and holding through volatility.
How to read this issue
- Verified: measured cost curves from primary research
- Interpretation: what those curves do to the real economy
- G.L.A Belief: our investing constitution, not a forecast you can take to the bank
- Coach’s Take: how we play it for the next 5-10 years, which we view as the Super Bowl of longterm investing.
What We Believe
G.L.A Rule #3 from our investing constitution is the playing field, or the “game board” for the next 5-10 years.
This innovation revolution is being powered by 5 platforms which rival the advent of electricity, the automobile and the telephone.
| Platform | What collapses | | — | — | | Artificial intelligence | The cost of intelligence and knowledge work | | Robotics and autonomy | The cost of physical work, driving, and delivery | | Energy storage and distributed power | The cost of electrons for AI, mobility, and the grid | | Public blockchains | The cost of trust, settlement, and capital formation | | Genomics and multiomics | The cost of reading, writing, and understanding biology |
This revolution is akin to the early 1900’s: electricity, the automobile, and the telephone arriving at once. That cluster lifted global growth from a crawl to a walk. ARK INVEST calls this cycle The Great Acceleration. We agree.
For about 120 years, global real GDP has lived in a 2% to 3% lane. Our research indicates that the average was still slightly above 3% heading into 2025, with the IMF still modeling the old world toward about $150 trillion of global GDP by 2030. ARK’s own research case is different:real growth accelerating to more than 7%, with GDP pushed above $190 trillion by 2030. ARK INVEST’s Big Ideas Report 2026 puts the 2030 print at 7.3% versus an IMF 3.1%.
So the narrative that is built into the markets is that the world will grow at the same pace, 2-3% per year going into the end of the decade, we believe that the narrative is waaaaaay off!! If our research is right, the investors who bet on the growth of 7% will see huge outperformance.
G.L.A Belief: If growth stays at 2.5%, the world economy doubles in about 28 years. If it runs at 7%, it doubles in about 10. We built our whole strategy — concentration, the Gauntlet, DCA on weakness, no margin, no market timing — for that playing field.
The Unsexy Secret
Wright’s Law is the engine.
Theodore Wright showed in 1936 that we learn by doing. For every cumulative doubling of production, costs fall by a fairly stable percentage. Ark Invest uses that law because, measured through 2015, it beat Moore’s Law on price forecasts by about 40%.
Moore’s Law is change as a function of time. Wright’s Law is change as a function of units made. When production compounds, costs do not drip lower.They stair-step down.Then new markets open. Then production doubles again.
An example of this is that for more than four decades, solar-panel prices fell about 20% every time global cumulative capacity doubled. Module prices dropped from about $106 per watt to about $0.38 per watt. That is a 99.6% collapse. Computers and batteries same thing. Bicycles and coal plants do not.
Interpretation:Most of the world still prices innovation like a coal plant. Linear, Slow, Mean-reverting.The companies we want are the ones riding Wright’s Law so hard that yesterday’s “too expensive” becomes tomorrow’s biggest winners.
The Cost-Collapse Stack
This is the innovation boom in one table.
| What got cheaper | The measured drop | Why it matters | | — | — | — | | GPT-3.5-level AI inference | More than 280-fold, $20 to $0.07 per million tokens, Nov 2022 to Oct 2024 | Intelligence becomes a utility | | LLM inference at a fixed capability | Median 50x per year; 9x to 900x depending on the task | Agents stop being a demo | | Utility-scale solar electricity | LCOE down 90% from 2010 to 2024, to $0.043/kWh | AI can buy cheap electrons | | Lithium-ion battery packs | 93% cheaper since 2010, to $108/kWh in 2025 | Storage, EVs, and the grid scale | | Stationary storage packs | $70/kWh in 2025, down 45% in one year | Data centers can time-shift power | | Human genome sequencing | About $95 million in 2001 to about $500 by 2022, a 190,000-fold drop | Biology becomes software | | Satellite bandwidth | 7,500-fold cheaper since 2004, $300 million to $40,000 per Gbps | Orbit becomes infrastructure |
Cheap intelligence, plus cheap electrons ,plus cheap launch, plus cheap biology = Sustainable Abundance For All.
The Cost of Intelligence Is Collapsing
Verified.Stanford’s 2025 AI Index found that querying a model at GPT-3.5-level MMLU performance fell from $20 per million tokens in November 2022 to $0.07 by October 2024._Hardware costs have been declining about 30% a year. Energy efficiency has been improving about 40% a year. The smallest model clearing a 60% MMLU bar shrank 142-fold in two years, from 540 billion parameters to 3.8 billion.
Epoch AI measured the same war from the other side. Across six benchmarks, the price to hit a fixed performance bar fell 9x to 900x per year, with a median of 50x. After January 2024 the median accelerated toward 200x. Epoch is honest about the caveat: the fastest drops are recent, so it is less clear they persist at that exact speed.
Our Research indicates that the same cost declines are happening inside the labs. AITraining costs falling about 75% a year. Inference costs falling about 90%.Roughly half hardware, half software. Training costs have been halving about every six months, versus the two years Moore’s Law made famous.
I can sum this up in one sentence-AI inference costs are collapsing, and that is driving rapid growth in tokens processed by developers, enterprises, and consumers.
Summaries of the ARK INVESTBig Ideas 2026report, add the 2025 tape:_inference costs down more than 99% in a year, OpenRouter-style demand up 25-fold after late 2024, software-development token costs down 91% in eight months, and AI agents’ reliable task duration up 5x in 2025, from 6 minutes to about 31.
Interpretation.This is why “AI spend is too high” is a 2022 sentence.Unit costs are falling faster than most CFOs can re-forecast. Total bills can still rise because usage is exploding. That is Wright’s Law doing its job. Solar did the same thing. Sequencers did the same thing. The market confuses a rising revenue line with a broken cost curve. They are often the same event.
Coach Kyle’s Take: We do not need to win a debate about which lab has the best demo. _We need the companies that sell the meters while intelligence goes from scarce to abundant:_compute, memory, cloud, and the software layer that puts agents inside real institutions. (NVDA MU CBRS PLTR etc)
The Cost of Electrons Is Still Falling
AI does not run on press releases. It runs on watts.
Verified.IRENA_says 91% of new renewable projects commissioned in 2024 were cheaper than the cheapest new fossil alternative.Utility-scale solar LCOE fell 90% from 2010 to 2024, to $0.043 per kWh, 41% cheaper than the lowest-cost fossil option. Onshore wind sat at $0.034 per kWh, 53% cheaper. Total installed solar cost fell 87%, from $5,283/kW to $691/kW. Utility-scale battery systems fell 93% from 2010 to 2024, from $2,571/kWh to $192/kWh.
2025 pack survey: average lithium-ion packs at a record-low $108/kWh, down 8% even as metal prices rose, and 93% below 2010. BEV packs at $99/kWh, under $100 for the second year. Stationary-storage packs at $70/kWh, down 45% in one year. China at $84/kWh. LFP at $81. Four-hour storage project costs hit a record low in BNEF’s LCOE work.
Let me be blunt:energy is becoming the binding constraint, and falling power and storage costs are how the next growth wave gets financed. The ARK INVEST Big Ideas 2026 report says cumulative global power investment needs to scale about 2x to roughly $10 trillion by 2030, with stationary storage scaling about 19x.
Interpretation. The bottleneck is not “is solar cheap.” Solar is cheap. The bottleneck is building the watts, the interconnects, and the storage next to the GPUs. That is why energy is no longer a side theme in this book.It is AI infrastructure.
Coach’s Kyle’s Take:TSLA is batteries, solar, Megapack, and the robot that will eat the cheap electrons. OKLO and BE are the nuclear and industrial-power surround. IREN is power turned into compute. Crusoe is the private version of the same idea. We are not trading power prices. We are owning the platforms that make intelligence cheaper to run.
The Cost of Atoms, Orbit, and Delivery
Reusable rockets pushed space onto an exponential cost curve. Again Innovation Booms are all about the cost declines.
Verified. Ark Invests research on wrights law says it all. Satellite bandwidth costs have dropped 7,500-fold since 2004, from $300 million per Gbps to $40,000. ARK’s learning rate: about 45% cost decline for every cumulative doubling of gigabits per second in orbit. With Starship, ARK’s case is another 40-fold drop toward about $1,000/Gbps by 2028.
The logistics math is simple.Last-mile app delivery around $15 versus drone and robot delivery under $1, about 90%. Truckload cost per ton-mile could fall about 60% as autonomous electric trucks replace human-driven diesel.
Ron Baron walked the launch curve on air last week: from tens of thousands of dollars per kilogram to a Starship world measured in hundreds. We covered that in the last Infinite Innovation issue. The point does not change. Cheap launch is the unlock for Starlink, direct-to-cell, and orbital compute.
Interpretation._Robotics is Wright’s Law in steel. _Every mile driven, every package dropped, every satellite flown teaches the system. Tesla FSD, Optimus, warehouse robots, maritime autonomy, and reusable launch are the same idea in different costumes: software that moves through the physical world, getting cheaper as the miles compound.
Coach’s Kyle’s Take: TSLA first. SPCX as the off-planet railroad. AMZN for warehouse robotics and logistics. ASTS as the space-to-phone satellite. Apptronik, Anduril, Saronic, and Shield AI on the private side. We do not need every robot in every kitchen in 2027. We need the cost curve to keep gliding so the TAM can show up.
Why Demand Explodes
Here is the part the tape never prices correctly.
When solar got 20% cheaper per doubling, the world did not buy 20% more panels. It found rooftops, deserts, data centers, and entire national grids that could not have used solar at $106 a watt. When genomes went from $95 million to $500, medicine did not get 190,000 times more “efficient.” It became possible. When inference goes from $20 to $0.07, you do not get a slightly cheaper chatbot. You get agents that can sit inside every workflow.
G.L.A Belief:The same wrong narrative exists today as it did many times in the past when innovation was ready for prime time—- Consensus still lives in a 3% GDP world._The innovation stack is already behaving like a different economy. Software spend, can accelerate because companies buy software instead of hours. _Robotaxis, household humanoids, and autonomous logistics only look “too early” if you freeze today’s unit costs. We will not freeze them.
How G.L.A Plays a Cost-Curve Boom
If you could harvest this by trading it, everyone would. They cannot.
Rule #1:short-term markets are a voting machine. Long-term markets are a weighing machine. Voting machines are popularity contests, Weighing machines measure a company’s true value with fundamentals.
Rule #2: We like to put all of our eggs in just a few baskets, then we watch those baskets very closely. Research, buy, hold, DCA on weakness, manage. Trim a slice after a violent run if research says recycle. Park it in STRC or another income-oriented asset, recognizing capital risk. Add to the name that is down with the thesis intact. Only sell for two reasons only: a better long-term home, or a broken thesis.
Rule #4 from the G.L.A BIBLE -know what we own. Six to ten diversified tech juggernauts, not six versions of the same bet. Visionary leadership. Companies that make Must-have products and services. Huge TAMs (total addressable markets), Moats that get wider as the cost curve glides. Companies that Wall Street is confused about.
The formula does not change:
Dollar-cost averaging + exponential growth + compounding = generational wealth.
We only invest in companies our research says can 10x, 20x, and 30x over five to ten years. We are extremely picky with what we let in our portfolio, most companies fail with what we are looking for. Knowing what not to do or what not to invest in is 90% of the battle.
What Can Go Wrong
Every thesis has its risks and things that can derail or delay.
The curves can flatten. Epoch already flags that the fastest inference drops may not persist at 200x a year. Solar LCOE ticked up 0.6% in 2024 even as installed costs kept falling. Tariffs, grids, and permitting can slow the glide.
Power can stay the bottleneck. Cheap models on an expensive grid is a science project. If interconnects slip, the boom becomes a long grind.
Adoption can lag the lab. Robotaxis, humanoids, and orbital compute can be right and still be late. ARK’s 7% GDP case assumes several S-curves fire together. That is the bull case, not the base case we require every name to clear.
Concentration cuts both ways. Six to ten names can 3x to 5x an index. They can also draw down 40% to 70% without the thesis dying. That’s why we don’t use margin, and that’s why we keep 40% of our equity in non correlated assets like gold, silver, bitcoin, STRC, and US treasuries.
KEY MAN risk is real. The operators we want are also single points of failure.
We mine those risks every day, we cannot predict when the next black swan event will happen but we can prepare for it.
G.L.A Bible Principle of the Week
Rule #3. Understand the game board before you pick the players.
Every element of the G.L.A strategy was built for this decade: AI, robotics, energy storage, blockchain, and genomics collapsing costs at the same time. If you underwrite 3% forever, you will own the index and wonder why the platforms look “expensive.” If you underwrite Wright’s Law, you will understand why the expensive thing is often the cheap thing on a five-year clock.
How To Use This Issue
When a headline says AI is a bubble, come back to the 280-fold inference raining price drop.
When a headline says energy is a tax, come back to solar down 90% and battery packs at $108/kWh.
When a headline says SpaceX is a science project, come back to bandwidth down 7,500-fold.
When a headline says Tesla is a car company, come back to the cost of physical work.
Then run the company through the G.L.A Gauntlet. If the company is still the company we think it is, hold. If it is down and the thesis is intact, add. If it is extended, trim a slice and wait for the next entry.
The early 1900s did not reward people who rented Edison for a week. It rewarded people who understood that electricity, cars, and phones were one boom, then owned the platforms long enough for compounding to do the violent part.
This is the innovation boom. It is all about cost declines. That is what we believe. That is how we will attack!
Coach Kyle G.L.A INVE$T RE$EARCH
This newsletter is for education and research discussion only. Coach Kyle is not a broker or registered investment adviser. Nothing here is personalized investment advice, an offer to buy or sell any security or cryptocurrency, or a recommendation to copy any portfolio. Innovation platforms are volatile.