So why copper, and why now? I think S&P Global’s January report, “Copper in the Age of AI,” called this one early. It projects that global copper demand will climb 50% by 2040, reaching 42 million tons a year, with AI, data centers, and defense alone adding about 4 million tons of that. The problem is supply. Global copper production is expected to peak at 33 million tons in 2030, then start declining, leaving a shortfall of more than 10 million tons, 25% of projected demand, by 2040.

It takes an average of 17 years to get a new mine from discovery to production. Compare that to the 3 to 5 years it takes to build a data center, and you start to see how differently this industry runs on time. AI-assisted exploration and refining could eventually shorten that mine timeline, but for now it is still a grinding, decade-plus process at best.

None of this is actually new, by the way. EVs, renewable power, grid buildout, the broader energy transition from a fossil-fuel-based system to an electrified one, have been pushing copper demand higher for more than a decade.

Even in S&P’s own report, energy transition demand (15.7 million tons) dwarfs the combined AI, data center, and defense number (4 million tons). Honestly, 15.7 million tons is hard to picture just by reading the number, so maybe just file it under “a lot more.” Still, the reason AI and data centers are getting so much attention within that smaller 4 million ton slice is that three years ago, this demand was not even in the forecasting models. S&P Global’s head of energy transition consulting put it well: three years ago, AI and data centers were not even on the radar. That alone says something about how fast this industry is having to recalculate.

Think of copper’s role in a data center as its vascular system. It is in the wiring inside server racks, the pipes in the cooling systems, the transformer windings, and the transmission infrastructure running all the way from the power plant to the campus. Last year alone, 100 new data center projects were announced globally, worth $61 billion. Every one of those comes with a massive copper bill attached.

The US actually has plenty of copper in the ground, but not enough refining capacity. The reason is simple: it does not pencil out. Refining requires enormous capital investment, and once you factor in America’s high cost of materials, labor, and everything else, the economics just do not work.

So even when raw ore does get mined, much of the work turning it into high-purity copper still gets outsourced, mostly to China. That is exactly why VCs are putting money into refining and reprocessing startups instead of mining companies. If opening a new mine takes 17 years, reprocessing copper scrap that has already been mined, faster and cleaner, domestically, gets your capital back a lot quicker. Did I mention it is expensive? These days, top-tier VCs are basically paying money to buy back that time.

To be clear, this is more of a hypothesis on my part than a confirmed trend, still an early signal at best. What I think actually matters here is not the dollar figures but how much the investment frame around AI infrastructure keeps expanding. Last year, every conversation in this industry was about securing GPUs. Then it was memory. Then it was power shortages. It is starting to feel like a show entering its next season. Now that raw material processing has made the list too, it suggests the bottleneck in data center capex is not just chips anymore, it is spreading across the whole supply chain. I could not tell you exactly where the next traffic jam will hit, but one thing seems clear: the further up the chain the bottleneck moves (chips, then power, then raw materials), the more room VCs have to get involved.

There is a real possibility that aluminum eats into some of copper’s share. It conducts less efficiently, but its weight and cost advantages already make it the choice for certain wiring segments, and that shift could accelerate if copper prices keep climbing.

And of course, there is a decent chance the headline data center numbers themselves are inflated. As Bloomberg pointed out in August, a large share of the US data center projects that have applied for power will probably never actually get built, and that phantom demand is likely baked into copper demand estimates too. So it is worth treating these copper demand projections with a healthy amount of skepticism rather than taking them at face value.

The actual scale of these refining startups, compared to established players, is still basically first-birthday-party stuff. There is a long way to go. Still Bright’s current throughput is about 2 tons a year. Against the multi-million-ton gap S&P is talking about, that is rounding-error territory.

Whether these startups actually close that gap, or end up as scale-up failure stories a few years from now, is not something I can call with confidence yet. But the fact that this much capital is already flowing into plants instead of mines is reason enough to keep watching where the next bottleneck breaks.

Sources

• Substantial Shortfall in Copper Supply Widens (S&P Global press release, 2026-01-08)

• AI to boost copper demand 50% by 2040 (MINING.COM, 2026-01-08)

• The future depends on copper… systemic risk (Fortune, 2026-01-09)

• AI’s appetite for copper poses a ‘systemic risk’ (SDxCentral)

• American copper processing sees renewed startup/investor interest (Axios, 2026-08-14)

• Red Metals attracts $10M investment for refinery (Recycling Today)

• AI data center boom, power grids strengthen investment case for copper (Bloomberg, 2026-08-28)

• Most Power Sought for US Data Centers Will Never Materialize (Bloomberg, 2026-08-12)

If grid access delays and local opposition are the two defining problems for data centers on the ground, neither one exists in space. No grid means no interconnection queue, and there are no neighbors next to a satellite to organize a protest. The US currently has more than 25 gigawatts of data center capacity under construction, and the pace of new grid interconnection approvals can’t keep up with that volume, which is the real bottleneck this industry is running into right now. It’s not surprising that going to orbit is being pitched as a way around it.

The company that stands out here is Starcloud. When I first read that they’d flown a single Nvidia H100 on a 60kg satellite and trained a language model in orbit back in November 2025, I assumed it was little more than a demo. Then the whole space started moving. In January 2026, SpaceX filed with the FCC for a computing-satellite constellation of up to a million units. Two months later at GTC in March, Nvidia unveiled its Space-1 Vera Rubin module and declared orbital computing an official product category. Google is preparing to launch a prototype for its Suncatcher project in early 2027. And Starcloud came back into view at the end of that run: on August 21st they raised another $250 million in a Series A extension, bringing their total 2026 raise to $420 million and their valuation to $2.3 billion. Their Series A valuation back in March was $1.1 billion, so that’s more than double in five months. I’ve been looking at venture deals for years, and that pace still catches me off guard every time.

The logic behind going to orbit, or more precisely to low Earth orbit (LEO), starts with physics. In space, without atmospheric attenuation, solar intensity holds steady at 1,361 watts per square meter, and a dawn-dusk sun-synchronous orbit keeps a satellite almost entirely out of Earth’s shadow. Ground-based solar capacity factors in the US top out around 25% because the sun rises and sets, but in LEO you get nearly uninterrupted sunlight around the clock. That’s the premise the whole industry is built on: that this can function as baseload power.

But nothing is ever really free, and this premise comes with new costs that don’t exist on the ground. There’s mass cost, since every kilogram of hardware translates directly into launch cost, and there’s thermal cost, since there’s no convection or conduction in a vacuum, so all heat has to be shed by radiation. Radiation turns out to be the more physically stubborn problem of the two.

The Stefan-Boltzmann law says the heat you can shed by radiation scales with the fourth power of absolute temperature (so I’m told), and at 50°C, radiating away 1 megawatt requires roughly 1,200 square meters of radiator panel, about a third of a soccer field. Since radiator area scales with computing power, a lot of people assume there’s a hard ceiling on how much of this you can put in orbit, or that costs balloon past a certain point.

Looking at that number alone, a large orbital data center seems physically impossible. But the systems actually being designed today are a different scale. Starcloud-2’s GPU is built to dissipate 7kW, and the radiator needed at that scale is only 8.4 square meters. The reason this works isn’t that anyone solved the megawatt-scale thermal problem. It’s that the current designs simply avoid that scale altogether.

Radiation exposure follows a similar pattern. Commercial chips can suffer bit flips or permanent damage from cosmic radiation, and the industry has shifted away from radiation-hardened chips that lag five or ten generations behind, toward off-the-shelf commercial chips paired with shielding and software-level error correction instead. Inter-satellite communication involves the same kind of tradeoff. Laser links are theoretically capable of terabit speeds, but they still fall short of what’s needed for large-scale training, which requires hundreds or thousands of GPUs tightly coupled together. That’s why what this industry is selling right now is inference, not training.

Let’s look at the cost side too. According to Gartner, space-grade solar panels cost 1,000 times more than terrestrial ones, orbital satellites have to survive thermal cycling between -173°C and 127°C (100K to 400K), and there’s no way to send a repair crew if something breaks. Even accounting for how expensive the space industry already is, a 1,000x figure is on a different level entirely.

Short-seller Jim Chanos calls it “AI snake oil,” arguing that once you add up launch costs, insurance, redundancy, and radiation-hardening, the total ends up dwarfing whatever you save on terrestrial electricity bills.

Even the industry itself sounds unsure about the cooling problem. At GTC 2026, Nvidia CEO Jensen Huang admitted: “In space there’s no convection, only radiation. How to cool this is still being researched.”

TechCrunch’s February 2026 analysis puts a sharper number on it. A 1-gigawatt orbital data center costs nearly 3x as much overall as an equivalent ground-based facility ($42.4 billion), and the gap widens further when you isolate power alone: annual cost per kilowatt runs $570 to $3,000 on the ground versus up to $14,700 in orbit. At that point, the skepticism starts to feel less like a fringe view and more like the consensus.

But read the skeptics’ language closely and a pattern shows up. Words like “right now,” “within this decade,” and “at scale” are all hedges. They’re not saying it will never work, just that it doesn’t work yet, and that gap is exactly what’s drawing capital in today. There’s already a place where this pencils out: processing satellite imagery, especially SAR (synthetic aperture radar) data, takes about two hours on the ground but only minutes in orbit. For workloads like counterterrorism operations or disaster response, where speed itself carries value, the cost premium stops mattering.

This isn’t a market competing head-on with terrestrial cloud for general-purpose AI training and inference. It’s a narrower, more specialized market where cash flows first. And companies like Aethero, Cosmic Shielding, and Kepler Communications, which sell radiation-hardened components or inter-satellite communication modules rather than complete services, are positioned to benefit regardless of how the broader orbital data center market plays out, as long as it keeps growing. Their bet is that you don’t need to pick the winner if you’re the one selling the picks and shovels. The political bottleneck on the ground plays into this too. That $64 billion in blocked projects doesn’t look like it’s going to resolve easily, and that’s one more reason capital keeps buying into the space option.

To sum it up: the optimists seem focused on the fact that this technology “works” at all in such an extreme environment, while the skeptics are focused on cost efficiency. It sounds like the same conversation, but the two sides are really operating on different scales entirely. The startups in this space are talking about data centers in the thousands of kilowatts, just now reaching 1 megawatt, because that’s as far as the technology has gotten. Meanwhile, the pessimistic scientists and investors in this space are already comparing that to 100-megawatt terrestrial data centers, and concluding that space is tens to hundreds of times more expensive, can’t be cooled properly, and just won’t work. A 100-megawatt comparison, of all things. It makes me wonder if the pessimists are, in their own way, more optimistic than the optimists here, since assuming space data centers will ever compete at that scale is itself a fairly hopeful assumption. Or maybe that’s just me getting ahead of myself.

There’s an irony worth noting. What Starcloud emphasized in its August raise wasn’t technology, it was securing launch capacity. With SpaceX’s Falcon 9 program winding down by 2028, CEO Philip Johnston has said they need to book an enormous volume of launches right now. The entire industry’s economics rest on the story that launch costs will keep falling over time, but what’s actually happening at the moment is a scramble to lock down launch slots, not lower prices. Those are two different problems, and until Starship reaches a stable commercial launch cadence, the slot shortage might turn out to be the more immediate constraint.

Whether this ends up as a bubble the skeptics warned about, or quietly takes root starting with defense and intelligence workloads, is still an open question. The real launch of Starcloud-2 and its first commercial customers in orbit, Google Suncatcher’s 2027 prototype, and Starship’s actual launch cadence: those are the three things worth watching next.

Sources: Data Center Watch, “$64 billion of data center projects have been blocked or delayed” / The Register, “Orbital datacenters are a pie-in-the-sky idea: Gartner” (Feb 25, 2026) / Data Center Dynamics, “Plans for space data centers labelled ‘ridiculous,’ ‘AI Snake Oil,’ and ‘peak insanity’” / TechCrunch, “Starcloud raises $250 million for orbital data centers as launch options dry up” (Aug 21, 2026) / SiliconANGLE, “Starcloud raises $250M to build AI data centers in orbit” (Aug 21, 2026) / Payload Space, “Starcloud Raises $170M Series A at $1.1B Valuation” (March 2026) / TechCrunch, “Why the economics of orbital AI are so brutal” (Feb 11, 2026) / Mordor Intelligence, “Synthetic Aperture Radar (SAR) Market”

The US isn’t short on electricity in any aggregate sense. What’s scarce is capacity during the handful of hours each year when demand spikes all at once, a July heat wave, a January cold snap. Utilities have to build for those peak hours, then watch the same infrastructure sit half-idle the rest of the year. A data center that insists on guaranteed full power at any moment, including the worst of those peaks, is effectively asking a utility to size its whole system around that one facility’s worst case. Multiply that across every new gigawatt-scale project trying to get connected, and you get the multi-year interconnection queues everyone in this space complains about.

What took me longer to appreciate is why nobody fixed this sooner. It isn’t a technology gap. It’s an incentive gap.

Regulated utilities earn their return on the capital they build, not on the capacity they help someone else use more efficiently. A new substation or transmission line goes into the rate base and earns a guaranteed return for decades. A software layer that lets existing infrastructure serve more customers does not. So for years, the rational move for a utility has been to keep sizing for peak and let developers wait in line, even when the electrons technically exist. That’s a structural reason flexibility never got built, not a technical one.

That gap is what a new category of infrastructure software is now walking into: systems that sit between a data center and the grid and decide, in real time, how much power to draw and when. If a cluster can shed load for twenty minutes during a demand spike and make it up later, the utility no longer has to plan around that cluster’s theoretical peak. Do that across enough facilities and you’re not adding generation, you’re unlocking capacity that already existed but was never allocated to anyone.

None of this is conceptually new. Industrial demand response has existed for decades, aluminum smelters and steel mills getting paid to cut load during grid emergencies. What’s different with compute is the response time. A smelter needs hours of notice. A GPU cluster can modulate in seconds to minutes without touching the workloads running on it, which makes it a far more useful, and more valuable, flexibility resource than anything the demand response industry has worked with before.

Emerald AI is one company building this layer. It raised $150 million earlier this year at a reported $1 billion valuation. That round came on the back of pilot results in Phoenix and London, where GPU clusters reportedly shed load during grid stress events without materially slowing training or inference. I’m less interested in the specific number than in what it signals: institutional capital is now willing to price grid orchestration for data centers as its own category, distinct from power generation and distinct from data center construction.

The open question, and the one I find myself going back and forth on, is who ends up owning this layer long term. Hyperscalers already run some of the most sophisticated internal power management in the world, Google has published research on shifting compute load to follow carbon-free power availability for years now. A hyperscaler with that internal capability may simply build its own version rather than buy it. The more interesting buyers are probably neoclouds and colocation operators scaling fast without that internal expertise, which is a smaller market than the headline funding numbers suggest, but a real one.

None of this works unless utilities are willing to credit curtailable load in their planning, which means capacity markets like PJM and ERCOT actually changing how they treat flexible demand, and that moves slowly and unevenly by region. It also only works if the economics hold: the value of the freed-up capacity (faster interconnection, avoided curtailment fees) has to exceed what it costs to build and operate the orchestration layer itself. Pilot results are not the same thing as economics that hold up at scale.

Dominguez’s framing stuck with me because it points at where the real leverage is. The industry has spent two years treating this as a generation problem, more gigawatts, faster permits, more turbines. Timing is a cheaper problem to solve than supply, which is exactly why I think the coordination layer between compute and the grid ends up mattering more than it gets credit for right now.

The $62 Billion Problem Hiding in Plain Sight

Here’s a stat that should make every infrastructure investor pause: globally, we generate the fastest-growing waste stream in the form of electronic waste, yet only 22% gets recycled. The UN estimates we’re literally burying $62 billion worth of recoverable materials annually. That’s not just an environmental tragedy – it’s a massive market inefficiency screaming for disruption.

For hyperscale operators like Amazon, this isn’t theoretical. They’re running hundreds of AWS data centers globally, with server refresh cycles every 5-6 years. The traditional approach? What the industry calls “smash and grab” – essentially destroying equipment and selling it as scrap metal. It’s primitive, wasteful, and increasingly expensive.

Why Amazon Wrote a $10M Check to Molg

Amazon’s investment in Molg’s Series A wasn’t charity. It was strategic necessity wrapped in an ESG bow. The company committed to net-zero carbon emissions by 2040, then watched their carbon footprint increase 6% last year as AI demand exploded. Something had to give.

Molg’s solution is elegantly simple: deploy AI-powered robots that can intelligently assess, test, and disassemble electronic components. Instead of destroying everything, these systems identify what’s still functional, what can be refurbished, and what should be recycled. It’s precision demolition meets circular economy.

The Virginia-based startup has installed these systems at Amazon’s re:Cycle Reverse Logistics facility in Pennsylvania, where robots now handle servers and hard drives with surgical precision. The results speak to the business case: Amazon sourced 16% of its data center components from internal refurbished inventory last year, extending component lifecycles by 1-2 years.

The Investment Thesis Writes Itself

From a pure financial perspective, this is a no-brainer. You’re simultaneously reducing procurement costs (buying fewer new components), decreasing disposal expenses, and potentially creating new revenue streams from refurbished equipment sales. The unit economics improve across multiple vectors.

But the real opportunity lies in scalability. Molg is a 35-person company founded in 2021, currently producing robots at “dozens” scale. Yet they’re already expanding to Kentucky, Dublin, and Singapore with Amazon. The addressable market? Roughly 10,000 data centers globally, all facing identical challenges.

The competitive moat is surprisingly wide. Very few companies are building automated e-waste processing solutions at enterprise scale. Molg also secured a $6.5M Department of Energy grant to develop laptop disassembly capabilities, proving the technology’s adaptability across device categories.

Beyond the Green Washing

What excites me most about this investment isn’t the sustainability angle – it’s the fundamental business model innovation. Traditional e-waste processing is labor-intensive and geographically constrained. Molg’s approach is capital-intensive upfront but scales beautifully, with AI models that improve over time and apply across device types.

CEO Rob Rawson-Shanks frames it perfectly: “How do we redeploy this equipment at the highest value?” It’s asset optimization through intelligent automation, with environmental benefits as a valuable byproduct rather than the primary driver.

The Bigger Picture for VCs

This deal exemplifies a broader trend I’m tracking: the convergence of operational efficiency and environmental responsibility creating genuine investment opportunities. The old binary thinking – profit versus planet – is becoming obsolete as technology enables solutions that optimize for both.

For early-stage investors, the lesson is clear: look for companies solving real operational problems that happen to have positive environmental externalities. The sustainability narrative helps with enterprise sales and regulatory compliance, but the core value proposition must stand on pure economics.

Amazon’s willingness to invest and expand this partnership validates both the technology and the market timing. When a $760 billion company deploys capital to solve waste management challenges, it signals that circular economy solutions have moved from nice-to-have to business-critical.

The future of enterprise infrastructure isn’t just about building bigger, faster systems – it’s about building smarter, more sustainable ones. And for investors willing to look beyond traditional SaaS plays, that future is creating some compelling opportunities right now.

*The robots disassembling today’s servers might just be training the AI models that power tomorrow’s data centers. That’s not just good business – it’s beautiful irony.*

Subscribe now

You’re running a growing e-commerce business, selling everything from jewelry to vitamins across the United States. Life is good until you discover that in New York, a whole bagel is tax-exempt, but slice it in half and suddenly it’s “ready to eat” – boom, you owe sales tax. Welcome to the wonderfully absurd world of American tax law.

This exact scenario drove Sam Ross, a former Airbnb product manager turned serial entrepreneur, to build Numeral. And apparently, investors think his frustration is worth $350 million.

From Travel Dreams to Tax Nightmares

Ross conceived the idea while traveling the world in 2018, running several e-commerce brands from the road. Back then, the digital nomad lifestyle was still novel – remote work wasn’t the default it is today. He operated an online jewelry store and still maintains a vitamin DTC site, but the freedom came with a bureaucratic price: navigating the labyrinth of sales tax compliance.

Everything changed after a 2018 Supreme Court ruling that required online sellers to collect sales tax wherever their customers are located, regardless of physical presence. Suddenly, what used to be a California-only problem became a nightmare involving thousands of jurisdictions.

The Numbers Game That’s Not a Game

Here’s where things get mathematically insane: the U.S. has over 11,000 tax jurisdictions, each with unique rules and deadlines. To put that in perspective, there were over 400 sales tax rate changes in just the first half of 2025. That’s more than two changes every business day.

A company doing $5 million in sales typically owes sales tax in more than 30 states, meaning they need to comply with nearly 10,000 jurisdictions. For a finance team, that’s not complexity – that’s torture.

Enter the AI Tax Whisperer

Founded in early 2023 by Ross and Matt DuVall (CTO with engineering experience at Stripe and Notion), Numeral took a different approach. Instead of trying to simplify tax law – impossible – they decided to automate it entirely.

The timing was perfect. AI hallucination risks are low with tax rules because most regulations are deterministic and well-defined. While other AI applications struggle with uncertainty, tax compliance thrives on clear, if complex, rules.

The Funding Frenzy

The market clearly agrees. Numeral raised $35 million in Series B funding just six months after an $18 million Series A, reaching a $350 million valuation. The round was led by Mayfield, with participation from Benchmark, Y Combinator, and Mantis – the VC firm founded by The Chainsmokers.

Yes, the EDM duo is now backing tax compliance startups. If that doesn’t signal we’re in a unique market moment, nothing will.

The Global Gambit

What sets Numeral apart isn’t just automation – it’s ambition. The platform handles tax filings in markets like Tanzania and Kenya, supporting compliance across 60+ countries. While competitors focus on U.S. complexity, Numeral is betting on global-first businesses that sell everywhere from day one.

“We’re filing taxes for people in Tanzania and Kenya, in all these small markets where traditionally they would have had to use a CPA firm,” Ross explained. For digital businesses, this isn’t just convenient – it’s revolutionary.

The Results Speak

The growth metrics are impressive: 3.5x year-over-year revenue growth and over 2,000 clients. The platform has processed over $5 billion in transaction volume and filed 150,000 sales tax returns. Client roster includes Ridge, Eight Sleep, Graza, Grüns, and Manus – brands that represent the new generation of digitally native businesses.

The Competitive Landscape

Numeral isn’t operating in a vacuum. Market leader Avalara, acquired for $8.4 billion in 2022, is preparing for re-IPO. Competitors include Anrok, Zamp, and others, all chasing the same frustrated finance teams.

But here’s the thing about tax compliance: it’s not sexy, but it’s absolutely essential. And in a world where finance teams are drowning in regulatory complexity, automation isn’t just nice to have – it’s survival.

The Investment Reality Check

The $350 million valuation raises eyebrows, but the fundamentals support it. Tax compliance is a subscription business with high switching costs and expanding market dynamics. As businesses grow, they need more jurisdictions, creating natural revenue expansion.

The platform offers error-free filing guarantees, shifting liability from customers to Numeral – a bold move that demonstrates confidence in the technology.

What This Really Means

Numeral represents something larger than tax automation. It’s AI finding its practical application in unglamorous but essential business processes. While everyone else builds chatbots and art generators, Numeral is solving real problems that cost real money.

“As operators, we lived the pain of sales tax firsthand: hours lost, filings missed, and constant second-guessing. We built Numeral to eliminate that friction entirely,” Ross noted.

The Bottom Line

In a market flooded with AI hype, Numeral is refreshingly practical. They’re not promising to revolutionize creativity or consciousness – just to handle the mind-numbing complexity of tax compliance so businesses can focus on what matters.

And if that’s worth $350 million, maybe the future of AI isn’t as complicated as we thought. Sometimes the best solutions are the ones that make boring problems disappear.

The bagel tax might still be absurd, but at least now there’s an AI that remembers whether it’s sliced.

Subscribe now

There’s a new startup gaining traction in the AI industry, branding itself as the "Netflix of AI." It's called Fable. The name might evoke a wistful tale or a bedtime story, but make no mistake—this is a bold, ambitious tech startup. They recently secured backing from Amazon’s Alexa Fund and are quickly making headlines with a generative AI platform that can produce entire animated videos based solely on prompts.

But recently, Fable announced a project so eccentric and audacious that it stopped many industry watchers in their tracks: using AI to reconstruct the missing 43 minutes of Orson Welles’ legendary film, The Magnificent Ambersons (1942). And here’s the kicker: they’re doing it without owning the rights to the film—or any formal approval from Welles’ estate.

Fable: From Cartoon Studio to Cinematic Resurrection

Fable initially built its reputation as an AI content engine—think of it as a “creative Lego toolkit,” letting users produce animated stories with AI support, even generating unauthorized episodes in the style of South Park as experimental showcases. The Ambersons project, however, marks something of a pivot. It serves as a technology demonstration for Fable’s new large language and multimodal models, capable of generating sophisticated narratives.

Filmmaker Brian Rose, who has spent the last five years working on digital restoration, is leading this initiative. The technological approach includes:

1) blending AI with live-action footage—casting contemporary actors to reenact missing scenes and digitally replacing their faces with those of the original cast,

2) reconstructing iconic sequences—such as Welles’ famous 4-minute tracking shot—with painstaking precision,

3) proceeding without securing the copyright or even attempting to notify the Welles estate.

Reading all this, one can’t help but sigh. While “reviving dead art with AI” sounds noble in theory, in practice we may be left with a digital puppet show—a simulation that strips away the creative nuances Welles once infused into every frame: his framing, rhythm, tonal direction, and the subtle inflections of his actors. The Welles family, surprisingly, wasn’t outraged by the act itself, but rather by the complete lack of communication. In fact, Welles’ daughter has already ventured into the commercialization of AI voice models, suggesting the debate isn’t simply a matter of “AI = evil.”

Romance or Illusion? The Limits of AI Artistic Restoration

To give credit where it’s due, Fable’s endeavor does have its allure. In Korea, for instance, one might imagine lost footage or rare archival materials being reconstructed with AI—possibly giving closure to long-time fans and historians. In music, we've already seen similar experiments: AI-generated "new" songs by The Beatles or synthetic recreations of Frank Sinatra’s voice.

But these efforts, however technologically impressive, can never fully recapture an artist’s original brilliance. Physical constraints—like the acoustics of a recording space, the accidental shake of a hand-held camera, the subtle squint of an actor’s eye—are part of what gives authentic art its irreplaceable character. The irony of reviving Ambersons is that its status as “the greatest lost masterpiece in film history” is rooted in these very gaps. It is the impossibility of completion that has given it mythic status. No matter how refined, what Fable is creating isn’t Welles’ Magnificent Ambersons—it’s an AI-generated reinterpretation.

Industry Implications: The Collision of IP, AI, and Creativity

This also raises pressing questions about IP and authorship. The paradox of Fable’s approach is this: in a future where every IP becomes just another training dataset for generative models, the unique voice of the creator risks being drowned out. Control shifts from storytellers to platform giants—Amazon, Google, and their ilk.

This case is a litmus test for how technological innovation will continue to blur the boundaries between creation and replication. Fable’s project sparks not only fascination but a deeper debate: how far is too far when it comes to recreating—or fabricating—the legacy of artistic works with AI? As generative models grow more potent, these questions will only gain urgency, especially in a world where Hollywood and Korea’s content wars are increasingly intertwined.

The Future of Art, or Just Another Disruption?

We’re entering an era in which the ghosts of art may be resurrected by machines—and that demands a reevaluation of what creativity truly means. Where is the line between homage and appropriation? Between restoration and reinvention?

Let’s be very clear: the lost 43 minutes of Ambersons will always belong to Welles. The 43 minutes that AI reconstructs, however, do not. This may become the defining landscape of our time—a world where human tragedy, technological flourish, and enduring questions about originality all coexist, uneasily, within the frame.

As an investor and observer of the AI and creative tech space, I can’t help but watch with cautious optimism. But I’m also reminded that while technology can enhance art, it should never pretend to replace the soul that made it real in the first place.

These days, funding rounds for AI startups feel like a rollercoaster—huge, volatile narratives swing between unicorn-scale Series E deals and early-stage teams still struggling to close a Series A. But in the middle of that chaos, something surprising happened in July 2025.

The mere mention of “LegalOn” was enough to pull major investors to the table—no pitch decks, no roadshows needed. SoftBank Vision Fund, Goldman Sachs, Mizuho Bank, and even Mori Hamada & Matsumoto (one of Japan’s Big Four law firms) all jumped in. If you just look at the cap table, it’s like an All-Star roster of the financial world.

So, why the frenzy?

Everyone knows AI is coming for the legal space—contract review, compliance, due diligence, you name it. But LegalOn isn’t just another startup chasing this trend. It’s one of the rare companies that actually makes AI “legal-grade”—compliant, contextual, and painfully detailed. And more impressively, it’s a homegrown Asian startup from Tokyo that's expanding fast into the US and Europe.

This begs the question: What is LegalOn, really? And why is now the time to pay attention?

Let’s dig in.

LegalOn = Lawyer-Friendly AI

At its core, LegalOn is known for a product called Review—a contract analysis platform powered by AI, but built with legal precision. The name is unambiguous, and the functionality is even more so. Unlike generic LLM-powered tools that return vague warnings—("Here’s something that might be risky")—LegalOn tells you exactly what to change and why, based on your corporate compliance standards and real-life best practices authored by experienced lawyers.

That level of precision is only possible because of one key differentiator: LegalOn is powered by over 50 bespoke “lawyer playbooks.” These aren’t just scripted prompts—they’re operational checklists that in-house counsel actually use, encoded into the model. And it makes sense. Both co-founders, Nozomu Tsunoda and Masataka Ogasawara, come from corporate legal backgrounds. LegalOn isn’t so much a product as it is a digital extension of their years of legal experience.

They’ve also nailed the automation angle. LegalOn can reduce review times by up to 85%, and—crucially—it’s designed so that legal teams can start using it from Day 1. That’s not easy in B2B SaaS. Most enterprise legal teams are terrified of shelfware. But LegalOn can drop into an on-premise-style workflow without disrupting anything.

The results speak volumes: In Japan, over 25% of public institutions already use it. Globally, more than 7,000 organizations are on board. In just the past year, revenue in the UK and US has grown 4x.

Beyond Contracts: AI Agents for Legal Ops

With its recently closed $50M Series C, LegalOn is looking far beyond the legal memo. It’s not content being “just” a contract reviewer.

Enter: Matter Management.

This new feature tackles the workflow before and after a legal document even enters the pipeline—tracking contract requests, assigning internal owners, linking documents together—all within the same interface.

And then came the masterstroke.

LegalOn announced a non-equity strategic partnership with OpenAI. This is not just an API hook-up. Their engineers are working directly with OpenAI’s technical team, giving LegalOn privileged access to the latest GPT models and the inside track on implementation strategies.

They’re not just using GPT—they’re customizing it deeply for a legal domain. LegalOn might become the first AI tool truly optimized for legal teams, from input to interface to inference.

As CEO Daniel Lewis put it: “We’re not training our AI on generic internet data. We’re training it on law firm-grade playbooks written by practicing attorneys. That changes everything.”

If Harvey and Clio proved you could raise massive checks with “Generative AI meets vertical SaaS,” then LegalOn takes it a step further: LegalTech + Lawyer-founders + Japanese execution quality + global expansion mode. It’s a rare mix.

Human-First, AI-Assisted Law

But here's what really earns respect from legal professionals—and VCs.

LegalOn is unapologetically "lawyer-in-the-loop." It's not pretending to replace lawyers. It understands the deeply contextual, nuanced, and organizational nature of legal judgment. Instead, its AI drafts with speed and accuracy, but hands off the final call to a human.

That’s not a flaw—it’s the feature.

The legal industry doesn’t need a judge in the cloud. It needs an intelligent, compliant assistant. LegalOn delivers exactly that, and by shouldering the repetitive grunt work, allows legal professionals to focus on where their expertise actually matters—judgment, strategy, negotiation.

It’s a turning point. Legal teams using LegalOn aren’t just faster—they’re operationally sharper. Cost-effective. Scalable. Error-resilient.

And here’s the real takeaway for investors:

LegalOn is not just another doc automation startup.

It’s the archetype of what vertical SaaS in the AI era should look like:

- Deep industry expertise

- Customized large language models

- Founders who were operators

- Japan-based roots + global ambition

With over $200M already raised from names like SoftBank, Goldman, MUFG, and JAFCO, this is no longer a high-risk bet. Risk’s been priced in. Now the question becomes: how much of the US LegalTech market can they grab?

The AI moment in legal isn’t coming—it’s here. And LegalOn might be the first startup proving that AI can deliver not just speed, but full-stack operational improvement in one of the slowest-moving industries on earth.

This isn’t some greenfield idea. This is execution-first Legal AI.

And if I were a VC reading this?

I’d be dropping a link to this in the company Slack right now.

Subscribe now

As a VC investor, I spend my days hunting for startups that'll change everything. This corner of the internet is where I dump my thoughts on wild ideas and crazy opportunities that keep me up at night.

What I'm obsessed with:

• AI stuff: Machine learning magic, generative tools, whatever's breaking the internet this week

• Robotics: Bots doing human things, autonomous everything, the robot revolution (hopefully friendly)

• Data chaos: Platforms that make sense of mess, infrastructure that actually works, insights that matter

• Energy madness: Clean tech that doesn't suck, renewables going mainstream, batteries that last forever

Look, the coolest ideas happen when people actually talk to each other. Founders with impossible dreams, investors spotting weird patterns, random folks who just get it - everyone's welcome here.