Nazaré’s Year in Review, Part 2
Building the MACHA Stack
In 2024, investor enthusiasm still overwhelmingly rewarded throughput, volume, and lock-in – the core heuristics of a cycle dominated by centralized compute and large-model scaling. These were the right instincts for the conditions that prevailed, and scaling remains essential to the evolution of AI. But as constraints compound and architecture matures, those heuristics begin to feel less predictive, and at Nazaré we have a different hypothesis.
For one, economic participation in scale has become increasingly capital-intensive. Access to NVIDIA (the dominant proxy) is already saturated, as reflected in its recent $4 trillion valuation. What’s more, participating in multi-billion AI megarounds like SSI or Thinking Machines is more like pursuing different forms of God-like AGI than profitability or return on invested capital.
Secondarily, but perhaps more importantly, AI’s “next wave” centers not just on the expansion of frontier-scale training, but on the emergence of systems designed around real-world limitations: cost, energy, latency, privacy, coordination, and physical deployment.
These architectures – which we refer to as MACHA – prioritize efficiency, modularity, and verifiability. They are materially easier and more profitable to build, adopt, and invest in. New primitives spanning infrastructure, execution, and trust are already reshaping how intelligence is produced, distributed, and secured.
The companies in our portfolio reflect this orientation and provide practical evidence that constraint-aware innovation is gaining traction. Their innovations stand on their own, delivering efficiency, trust, and adaptability in environments where scale is limited or prohibitively expensive.
At the same time, these same innovations enhance the utility of whatever the prevailing state of the art happens to be. As large-scale models continue to advance, the tools that make them more efficient, deployable, and verifiable become even more valuable.
As such, this dynamic positions our portcos to benefit in two dimensions:
From the success of their own constraint-aware approaches
From the compounding improvements of the broader ecosystem (especially if scale does continue to deliver increased performance)
What follows is a reading of the terrain, grounded in the companies we’ve backed so far. Each investment contributes to the emerging architecture, and together they illustrate the shift we believe is already underway.
Compute
At the base of the stack lies compute. In 2024, compute was treated as a monolith – expensive, centralized, and largely provisioned through hyperscaler APIs. The assumption was that access to AI meant access to proprietary infrastructure. This is the Sun Microsystems mistake. If the architecture of intelligence is shifting from monolithic to modular, then compute must follow suit. It must become more available, more efficient, and more composable.
We saw this materializing early in Vast.ai aggregating underutilized GPUs into a global marketplace. They’re already profitable, with more than 175,000 GPUs online and tens of thousands of DAUs. Their early success confirms that the cloud isn’t the only place to find compute. Given the right abstraction, distributed infrastructure can deliver on performance, predictability, and price, providing reliable, scalable compute alongside traditional cloud environments.
Aion extended this insight to the enterprise layer. By serving the market for high-performance compute and streamlining payments through stablecoins, Aion is building a cloud service that mirrors the reliability of traditional providers while tapping into an entirely different supply base and consumer pool. The business speaks in the language of service-level agreements and procurement cycles, but underneath, it is composability and decentralization that do the heavy lifting. In a market shaped by rising cost pressures and volatile supply chains, architectural flexibility directly influences competitive durability and deployment scope.
CoopHive, on the other hand, treats compute not as infrastructure to be rented but as a resource to be traded by agents. It introduces a negotiation layer – a set of primitives for bundling assets, managing commitments, and enabling resource coordination among autonomous systems. The implications go beyond simple GPU access, laying the groundwork for agents to make economic decisions in real time across compute, bandwidth, storage, and data.
It introduces infrastructure that enables workloads to be negotiated dynamically, reflecting the changing requirements of multi-agent coordination and real-time resource exchange.
Intelligent Internet takes a broader view. Its architecture prioritizes openness, public good alignment, and edge deployment. By treating intelligence as a shared utility rather than a proprietary asset, it proposes an alternative trajectory – one in which accessibility, transparency, and customization become defining traits of global AI systems.
Training
Training remains critical, but it, too, is evolving. Not every model must be trained from scratch, but those that are must contend with new constraints: limited data access, variable compute availability, and the growing importance of adaptability. Centralized, brute-force training pipelines struggle in this context. We saw an opportunity for a different approach.
Prime Intellect reimagines model training as a dynamic, decentralized process. It combines adaptive optimization algorithms with collaborative training infrastructure to make model development more efficient, resilient, and open. The team’s background reflects deep experience with both foundational model architectures and distributed science. Prime Intellect’s design accommodates decentralized development and dynamic feedback. Training thus becomes less about monolithic runs and more about adaptive, composable cycles suited to distributed compute and real-world constraints.
Trust & Verification
As compute and training become more accessible, a parallel need emerges: proof. Inference is scaling faster than verification, and outputs are taken at face value because we lack the tools to question them. When autonomous agents, financial systems, and compliance workflows begin to rely on model decisions, however, the system must be able to justify itself.
Provably builds infrastructure for verifiable computation, integrating with existing data systems to provide auditability and integrity at the query layer. It permits existing databases to produce cryptographic proofs about their data queries and transformations without migrating the data or introducing specialized languages. It produces integrity guarantees at the level where most systems actually operate. In a world where compliance frameworks grow stricter and data lineage becomes legally material, this kind of infrastructure becomes necessary very quickly.
Hellas brings similar rigor to inference itself. Its architecture supports provable model execution: every output can be audited, verified, and reproduced. This becomes essential when an agent consumes the result of a model and acts on it. Hellas proofs serve as coordination tools within multi-agent systems, providing composability guarantees and enabling enforcement of execution policies without requiring implicit trust.
Verification also implicitly involves benchmarking. Layerlens provides real-time, transparent evaluations of open models across datasets, tasks, and hardware – enabling developers, researchers, and enterprises to make decisions grounded in reproducible data. While performance claims have historically lived in carefully-curated presentations and online leaderboards, LayerLens treats model evaluation as infrastructure: generating empirical, reproducible benchmarks that support decision-making.
Agents
As the systems that run AI evolve, so too do the systems that compose and direct it. The shift toward agent-based software changed the requirements for infrastructure because agents differ from conventional applications. They operate autonomously, maintain persistent state, and respond to evolving conditions across extended time horizons. They operate across time, not just in response to requests, queries, or inputs. Building for them requires new assumptions about identity, delegation, memory, and execution context.
Xoul abstracts this complexity. It’s a no-code interface and underlying protocol for creating autonomous agents with persistent memory and programmable behavior. More importantly, it treats agents as modular artifacts – created, reused, bought, sold, or shared. As an eventual agent “marketplace” Xoul turns the agent economy from an experimental playground into a composable layer for real-world tasks. As teams and individuals begin to deploy specified agents for sales, operations, development, research, and more, Xoul serves the clear need for structured creation and governance.
Privacy & Encryption
Finally, we return to privacy – a foundational requirement for many of the most promising applications of AI. Domains such as healthcare, personal data, and institutional knowledge depend on systems that maintain confidentiality and meet stringent data protection standards. In these environments, privacy governs both technical architecture and product eligibility. It defines the conditions under which intelligent systems can operate responsibly and gain long-term adoption.
Fairmath is an open-source research initiative building a modular library of privacy-preserving components using fully homomorphic encryption (FHE). Its FHERMA challenge platform crowdsources efficient methods for computing on encrypted data – including machine learning and analytics – without decryption. Winning solutions are added to a public component library, lowering the barrier for developers building privacy-first systems across AI, blockchain, and sensitive enterprise domains. Fairmath reflects a broader truth embedded in our thesis: that open-source coordination often outpaces closed systems in solving complex, constraint-bound problems.
Last Word
Each of these companies advances a specific component of the Nazaré thesis. Their work collectively demonstrates the emergence of a new architectural pattern across the AI infrastructure landscape: infrastructure and systems that do not depend on continued improvements in frontier-scale performance to generate value.
They validate the idea that the next wave of AI infrastructure will be built for efficiency, composability, and verifiability. More importantly, they suggest that this wave is already in progress.
The architecture is already taking shape, and because the market has yet to reprice around it, the most important parts of this new stack remain both investable and dramatically undervalued.
7@nazare.io
🤖🌊