Solivera Research
Deep technology

Research note

The bottleneck economy

For two decades, software set the pace of technology. The frontier is now moving back into the physical world, and the binding constraint is shifting down-stack.

Solivera Capital 8 min read
A funnel of layers narrowing downward Stacked bars narrowing from a wide abstraction layer at the top to a constrained physical layer at the bottom. SOFTWARE · DISTRIBUTION · DATA compute · silicon the physical layer ↑ ↓
The constraint moves down-stack. As systems scale, the binding limit shifts from code toward compute, power, materials and sensing.

The argument in three lines

  • Software set the pace of the last two decades, low marginal cost, fast iteration, global distribution, and it remains central. What is changing is that its next gains increasingly depend on physical systems that cannot be abstracted away.
  • The frontier, AI, robotics, advanced manufacturing, energy, is increasingly constrained by the physical layer: compute, power, chips, materials, sensing and deployment. The largest platforms are revealing this through their own capital allocation.
  • Solving for a bottleneck is not automatically a good business. What matters is where solving a physical constraint creates durable, company-level value, rather than a market that grows while it commoditises.

01 - The setup

The software era abstracted the physical layer away

For most of the last two decades, the economics of software set the pace. The most attractive companies scaled through code, distribution and network effects at near-zero marginal cost. Cloud computing abstracted away physical complexity, app stores reduced distribution friction, and open-source lowered engineering cost. This was a genuine structural advantage, and it produced some of the most valuable companies ever built.

What is changing is not the importance of software but where its next gains come from. Increasingly, software-enabled progress runs on physical systems that cannot be abstracted away: data centres, power, semiconductors, memory, cooling, materials and sensing. Companies built around those systems behave differently from pure software, they meet physics, manufacturing, qualification, reliability, regulation, supply chains and capital intensity. These are not shortcomings; they are the properties of building things that touch the physical world. As more of the frontier depends on them, they become a larger part of the story.

02 - The evidence

The frontier is moving down-stack

The clearest evidence is the behaviour of the largest technology companies. The leading AI and software platforms are no longer scaling only through code, data and distribution. They are securing compute, building data-centre capacity, designing custom silicon and controlling more of the physical infrastructure that growth now depends on.

OpenAI · Stargate
$500bn

Planned investment in AI infrastructure over four years, with $100bn intended for immediate deployment.2

Anthropic · AWS + Google
1M+chips

Trainium2 chips training and serving Claude through Project Rainier, alongside multi-gigawatt TPU capacity contracted with Google. Frontier AI as structured access to compute at scale.35

Amazon · Project Rainier
~500k online

Trainium2 chips already running in a purpose-built cluster. AWS is shifting from general-purpose cloud toward AI-specific silicon and infrastructure.4

Google · Ironwood TPU
42.5exaFLOPS

Per pod of 9,216 custom TPUs, built for inference-age workloads. Advantage moving into the integration of chips, systems and data centres.9

Meta · MTIA
$10sbn / yr

Custom training and inference silicon deployed at data-centre scale, across successive MTIA generations and tens of billions in annual infrastructure spend.1112

Tesla · Cortex
67,000H100-eq

Of training compute after adding 16,000 H200 GPUs in 2025. Autonomy and robotics treated as compute-and-sensing infrastructure problems.13

Apple is the qualitative version of the same shift. Apple Intelligence runs on the device where it can, and falls back to Private Cloud Compute, server hardware built around Apple silicon, for heavier requests, tying intelligence to a hardware-controlled privacy architecture.7 The bottleneck there is not compute capacity but trusted compute capacity.

Across very different businesses the pattern is consistent. A company known for models is now bound to one of the largest infrastructure projects in technology; a social-media company is designing its own chips; a cloud provider is rebuilding itself around AI-specific silicon. Constraint have moved.

Software advantage increasingly depends on physical capability.

03 - The system constraint

Everything routes through energy

The company-level signals are reinforced at the system level, and the system-level constraint is electricity. The International Energy Agency projects that data-centre electricity consumption will roughly double to around 945 TWh by 2030, growing about 15% a year from 2024, more than four times faster than electricity demand from all other sectors combined.1

AI does not consume electricity in the abstract. It consumes it in specific data centres, on specific grids, in specific locations, with real limits on power availability, cooling, transmission and permitting. Data-centre demand becomes a grid problem, a power problem, a real-estate problem and a cooling problem all at once. The constraint is not only how many GPUs exist, but whether the physical environment can support them, which is why power electronics, grid infrastructure, cooling and efficiency move up the list of things that matter as AI scales.

04 - The investment question

A bottleneck is not automatically a business

The bottleneck economy changes how companies are underwritten, but it is not a mandate to buy anything physical. The word itself can mislead: some bottlenecks are temporary, some are solved by incumbents, some commoditise, and some demand more capital than the value they capture. A technology can be genuinely important and still be a poor investment if it does not retain value.

The useful lens is value location. Four things tend to separate a strategic position from a merely important one: the bottleneck constrains a growing system rather than a niche; the solution becomes embedded in workflows, supply chains or infrastructure that are hard to replace; value accrues to the supplier rather than being captured by customers as the market grows; and the capability serves multiple downstream uses rather than one.

Timing

Deep-technology risk does not fall in a clean step; it changes form. Scientific risk can recede while engineering, integration, manufacturing and deployment risk remains, which is why a technology can be largely proven and still some distance from a finished business. Reading where a company sits on that path matters as much as the size of the market it addresses.

05 - Conclusion

The next constraints are physical

The next wave of technology will not be constrained only by code. It will be constrained by compute, energy, materials, sensing, manufacturing and reliability, not a single market, but a set of interdependent systems beneath the technologies the world wants to scale. The largest platforms are already revealing this through their own capital allocation: the advantage they are buying is increasingly physical.

That does not make every hardware or infrastructure company attractive. Some technologies will matter but commoditise; some will grow while supporting many winners; some will be impressive yet never become strategically important. Others will become embedded and increasingly valuable as their ecosystems mature. The interesting work is telling them apart, distinguishing the constraints that become durable, company-level advantages from those that are real but transient.

The future will still be shaped by our interactions with software, but the next constraints, and many of the next opportunities, are increasingly physical.