The thesis that made two decades of technology investing look smart was actually pretty simple: a lot of industries were running on processes that software could do better, faster, and cheaper. Find them, build the software layer, capture the value.
By any measure, that simple approach worked. Across financial services, logistics, media, healthcare administration, and retail, the gains were real and the market caps reflected it. Stripe built the payments layer. Shopify built the commerce layer. Uber built a dispatch layer over the existing transportation system. The pattern was consistent: identify an industry running on legacy processes, insert a software abstraction, and capture the margin that inefficiency had been sustaining. The TAM was always knowable because the underlying market already existed.
But this approach worked so well that it’s now largely played out. The industries most vulnerable to this kind of disruption have been disrupted.
What’s left is harder, and different in kind.
What’s Next?
The problems defining the current generation of frontier companies — in AI, energy, biology, robotics, and physical computing — don’t necessarily yield to better software layers. They require replacing underlying systems entirely. That’s a different class of bet, and the capital is starting to reflect it.
Of course the clearest signal of what’s next is coming from the balance sheets of the major AI platforms. Amazon has said it expects to invest about $200 billion in capital expenditures across Amazon in 2026, citing demand for “AI, chips, robotics, and low earth orbit satellites.” Alphabet raised its full-year 2026 capex guidance to $180 billion to $190 billion, driven by “unprecedented” demand for AI compute. Meta now expects 2026 capital expenditures of $125 billion to $145 billion, up from its prior $115 billion to $135 billion range, reflecting higher component pricing and additional data center costs. Microsoft reported capital expenditures of $34.9 billion, $37.5 billion, and $31.9 billion across the first three quarters of fiscal 2026, with much of that spending directed toward GPUs, CPUs, finance leases, and cloud infrastructure.
Rather than merely funding another software cycle, these companies are committing hundreds of billions of dollars to the physical substrate of AI.
Perhaps lost in the fervor over those unprecedented AI investment numbers is the fact that, by May 2026, more than $90.9 billion in venture and strategic capital has been deployed across six frontier sectors that in the past were considered moonshots: space, small modular reactors, humanoid robotics, quantum computing, synthetic biology, and brain-computer interfaces. Deep tech’s share of total VC funding has grown from roughly 10% a decade ago to about 20% today.
Optimization companies face execution risk: will this team ship the product, reach customers, and hit sustainable unit economics before the money runs out? Replacement-class companies face something harder to model: physics risk, long regulatory timelines, and the kind of structural uncertainty that comes with building categories rather than entering them. You can know within 18 months whether a logistics software company’s unit economics are working. A fusion company might execute perfectly on every engineering milestone for five years and still face open physics questions.
The Companies Making the Argument
SpaceX didn’t improve on existing launch economics at the margin. It demonstrated that the entire operating model of aerospace was wrong at the root. In 2025, the company completed 165 orbital launches, roughly half the global total and 85% of all U.S. missions. SpaceX’s reusable Falcon 9 costs approximately $74 million per launch. A single-use Atlas V, United Launch Alliance’s legacy heavy-duty launcher, costs roughly twice that amount.
Anthropic raised $30 billion in February 2026 at a $380 billion post-money valuation. Run-rate revenue at the time was approximately $14 billion, growing more than 10x annually for three consecutive years. By April 2026, Google had committed up to $40 billion and Amazon up to $25 billion in additional strategic investment in Anthropic. Both concluded that Anthropic’s alignment and interpretability research is the trust layer that enterprise AI deployment will run on.
Commonwealth Fusion Systems raised $863 million in August 2025, the largest deep tech and energy raise since CFS’s own $1.8 billion Series B in 2021. Construction of SPARC, its compact fusion demonstration machine, is well advanced at its Devens, Massachusetts campus. The U.S. Department of Energy has independently validated its full-scale magnet testing milestones, and CFS’s first commercial power plant in Virginia is targeting grid delivery in the early 2030s.
Helion Energy signed the world’s first power purchase agreement for fusion electricity in 2023, with a target of providing electricity to Microsoft by 2028. Construction on the Orion plant began in July 2025. Microsoft contracted for electricity before the technology was even commercially demonstrated.
Colossal Biosciences bet that synthetic biology had advanced far enough to do something conservation biology had never attempted: returning vanished species to functioning ecosystems at the genomic level. The de-extinction technologies behind that goal had never existed in the field before. But those technologies are now running programs for animals most people assumed were gone for good: the woolly mammoth, the dodo, the Tasmanian tiger. In April 2025, the company announced the successful revival of the dire wolf. The same platform is also being applied through conservation partnerships to endangered and at-risk species still living today. The category Colossal entered at founding didn’t have a market size, because the market didn’t exist. The applications of what it built are still expanding.
The frontier of that same synthetic biology category is also where Astromech is operating. Co-founded by Ben Lamm and Harvard geneticist George Church, Astromech is building predictive AI infrastructure for biology itself — using machine learning to forecast evolution, disease risk, and system-level vulnerabilities before they manifest. Rather than replacing a biological market that already exists, the company is creating the computational category that makes AI-driven biology navigable. It raised $40.5 million and reached a $2 billion valuation within nine months of launch — at the convergence of AI infrastructure and synthetic biology, the two replacement-class categories attracting the most capital in 2026.
What’s notable about these five cases is the breadth. Energy, AI infrastructure, fusion power, physical computing, genomics: these aren’t bets on a single breakthrough category. They’re evidence of a sector rotation playing out across the physical sciences simultaneously. The closest historical parallel is software itself in the 1990s, when capital followed a new class of infrastructure technology across every major industry at once.
The Investment Logic
Optimization companies can describe their addressable market at founding: the category exists, and the question is penetration. Replacement-class companies can’t do that. The category they’re creating has no direct historical comparable. Near-term uncertainty is higher. Over a long enough horizon, it’s substantially lower, because a company that succeeds at replacing an underlying system doesn’t compete for share of an existing market: it defines the terms of the new one.
The distribution of outcomes in venture is extreme, but within it, the outsized returns have often gone to the companies that looked most unlikely at founding. These are the companies making a moonshot-level commitment to replacing an existing system rather than optimizing it.
The investors who backed SpaceX when it looked like a stunt, CFS when commercial fusion had been “twenty years away” for half a century, and Anthropic when a safety-focused AI lab looked more like a research nonprofit than a business are now sitting on some of the most valuable positions in history. Companies that set out to make existing systems irrelevant generate returns that optimization companies, especially in the wake of the last 20 years of optimization, simply cannot.
Disclaimer: GeekWire newsroom and editorial staff were not involved in the creation of this content..
