The companies getting the most attention in AI right now are foundation model providers and the enterprise software layer built on top of them. Both categories are real. Neither is where the most interesting early-stage bets live. The foundation model market has already consolidated around a small number of players with enormous capital requirements. The enterprise software layer is increasingly competitive on every horizontal. What I am looking for is something different: AI companies with a physical data moat, embedded in consequential real-world infrastructure, solving problems where the competitive advantage compounds with every operational hour. Three companies fit that description.
Cradle — Amsterdam, Netherlands / Zurich, Switzerland | Founded 2021
What they do: Cradle builds an AI platform for protein engineering. Proteins are the molecular machines that execute virtually every biological function — the targets of most drugs, the catalysts of industrial processes, the structural components of advanced materials. Engineering them means modifying amino acid sequences to achieve specific properties: better stability, higher binding affinity, improved manufacturability, reduced immunogenicity. Cradle applies large language models to protein sequences, treating the amino acid alphabet the way GPT treats language, to predict which sequence modifications will produce desired properties and dramatically reduce the number of experimental rounds required to develop a functional protein. The company maintains a wet laboratory in Amsterdam specifically to validate its AI predictions against real biological outcomes and build the proprietary datasets that make its models better than anything trained on public data alone. CEO Stef van Grieken and the founding team built this from first principles at the intersection of machine learning and wet lab biology. The SaaS model with no royalties or IP complications is a deliberate structural simplification that removes the primary friction point in pharmaceutical AI partnerships.
Revenue: Not publicly disclosed. Enterprise contracts with Novo Nordisk, Johnson and Johnson Innovative Medicine, Novonesis, and Grifols at seven-figure annual values. SaaS model — no royalties, no IP complications.
Latest raise + backers: $73 million Series B in November 2024, led by IVP, with Index Ventures and Kindred Capital participating. Series A was $24 million in November 2023, led by Index Ventures. Total raised: $103 million. IVP is a growth-stage generalist VC. Index Ventures is known primarily for enterprise software, consumer, and fintech — not dedicated life sciences. For a protein engineering AI company with Novo Nordisk and J&J as customers, this investor composition represents a meaningful discovery gap in institutional biotech capital.
IP: Cradle holds patents on its AI-driven protein sequence prediction methodology, including language model architectures adapted for amino acid sequences and multi-property optimization algorithms. The proprietary wet lab dataset — built from A/B testing thousands of protein variants across multiple protein families — represents a training advantage that cannot be purchased on a market and cannot be approximated by fine-tuning on public databases such as UniProt or PDB. This proprietary experimental feedback loop is the specific moat that distinguishes Cradle from any team that fine-tunes a foundation model on public protein data.
Top customers: Novo Nordisk (the world’s leading GLP-1 drug developer). Johnson and Johnson Innovative Medicine. Novonesis (the world’s largest enzyme company, formerly Novozymes). Grifols (a major plasma-derived medicine company).
Moat and defensibility: Protein engineering AI has two compounding defensibility layers. The first is the quality of the training data: proprietary experimental measurements from wet lab validation are categorically superior to any fine-tuned model trained on public sequence databases, because they include the specific property data that determines whether a protein works in a real product. The second is the customer data loop: when Novo Nordisk runs protein optimization through Cradle, the outcomes improve Cradle’s models with each engagement. The model serving Novo Nordisk in 2026 is more capable than the one deployed in 2024 — not because of architecture changes alone, but because of what it has learned from real pharmaceutical protein engineering at industrial scale.
SWOT:
| Strengths | Weaknesses |
| Four marquee pharmaceutical and biotech customers representing the most demanding protein engineering organizations in the world. Proprietary wet lab dataset represents a non-replicable training advantage. SaaS model with no royalties eliminates IP negotiation friction that derails other AI-drug discovery partnerships. Amsterdam wet lab creates a validation feedback loop that pure-software competitors cannot access. | $103M raised puts Cradle at the upper end of what this article defines as modest funding. The path from AI tool to essential pharmaceutical infrastructure is long and dependent on continued customer success outcomes. |
| Opportunities | Threats |
| GLP-1 drug class expansion and broader biologics pipeline growth are driving unprecedented protein engineering demand. Industrial enzymes and sustainable chemistry represent a large parallel market. AI regulatory frameworks for pharmaceutical development tools are developing in a direction favorable to validated computational tools. | DeepMind’s AlphaFold successors are raising the capability floor for all protein structure prediction. Well-funded competitors including Absci and Recursion are building adjacent AI-biotech platforms. If foundation models commoditize protein AI generally, differentiation must come entirely from the proprietary dataset. |
Regulatory and compliance hurdles: Cradle’s platform is a scientific tool, not a drug, which means it does not require FDA or EMA approval. However, Cradle increasingly intersects with GMP-compliant data management requirements as it supports IND-enabling studies. EU AI Act provisions affecting AI used in high-risk applications — which includes pharmaceutical development — will require Cradle to maintain explainability and documentation standards for platform outputs. GDPR considerations apply to any customer data processed on EU-based servers.
Go-to-market: Cradle sells directly to the protein engineering teams at pharmaceutical, biotech, and industrial biology companies. The SaaS model eliminates the traditional friction in AI-drug discovery partnerships: no co-ownership of resulting IP, no royalty arrangements, no equity participations. This simplicity accelerates the procurement process significantly compared to competitors who propose co-development structures. The wet lab in Amsterdam serves a dual function: it validates Cradle’s predictions against real experimental data and serves as a demonstration asset for new customer acquisition.
GP Lens: Protein engineering is one of the most consequential industrial and pharmaceutical activities on the planet, and it has historically been a process of expensive, slow, partially random iteration. Cradle’s technology directly attacks the cost and time structure of that process. What gives me conviction is not the AI platform in isolation — it is the customer list. Novo Nordisk is the world’s most successful GLP-1 drug developer, with the most active protein engineering pipeline of any pharmaceutical company right now. The fact that Novo Nordisk chose Cradle as a platform partner means the technology performed in a real pharmaceutical engineering context. Grifols and Novonesis represent entirely different protein classes and application domains, which means the technology generalizes. That combination of validation depth and breadth is exactly what I am looking for when evaluating an AI company’s true moat.
Phaidra — Seattle, WA | Founded 2019
What they do: Phaidra builds Alfred, an AI agent that acts as a virtual plant operator for mission-critical industrial facilities — data centers, pharmaceutical manufacturing plants, and industrial facilities. Alfred connects to a facility’s existing building management system or SCADA infrastructure through standard protocols, continuously monitors thousands of sensor signals, and uses reinforcement learning to autonomously optimize cooling, power, and equipment settings. No new hardware required. The system trains in shadow mode before making live control decisions, and operators maintain full visibility and override authority at all times. Founded by Jim Gao, who led Google DeepMind’s industrial AI efforts and helped reduce energy consumption at Google’s data centers by 30%; Vedavyas Panneershelvam, a primary engineer on AlphaGo; and Katherine Hoffman, who led operations innovation at Trane Technologies and Raytheon. The team that solved this problem at Google went and built the same capability as a standalone company.
Revenue: Approximately $15 million ARR as of 2025, growing from $8.8 million in 2023. Multi-year contracts with hyperscale data center operators. STT GDC Singapore is a named customer.
Latest raise + backers: Approximately $110 million total raised. Series B: $50 million in October 2025, led by Collaborative Fund, with NVIDIA, Index Ventures, Sony Innovation Fund, and Helena participating. Mustafa Suleyman and Mark Cuban are individual investors. NVIDIA’s participation is both capital and strategic validation — Phaidra is embedded in NVIDIA’s Omniverse platform for new AI factory design.
IP: Filed patents on reinforcement learning architectures for industrial control systems, the digital modeling methodology that captures facility-specific behavioral physics before live deployment, and the multi-site orchestration framework for managing multiple facilities simultaneously. The facility-specific behavioral model is the core moat: it takes weeks to develop for a new site and improves continuously with operational data. Unlike static control logic that degrades over time and requires manual reprogramming, Alfred’s model gets more accurate with every hour of operation.
Top customers: NVIDIA (Omniverse AI factory design platform partner). STT GDC Singapore (data center operator). Pharmaceutical manufacturing customers under multi-year contracts (undisclosed by agreement). Multiple hyperscale data center operators under multi-year contracts (undisclosed by agreement).
Moat and defensibility: The facility-specific learning model is the moat. Every month of deployment produces behavioral data that makes Alfred more accurate for that specific facility — and that data cannot be transferred to a competing system without starting the learning process over from zero. The longer Alfred operates at a facility, the larger the performance gap between Alfred and any replacement that starts cold. NVIDIA’s Omniverse partnership opens the new-facility design market: AI factories embedding Phaidra before they go operational, establishing the behavioral model from day one.
SWOT:
| Strengths | Weaknesses |
| Google DeepMind AlphaGo engineer and industrial AI team as founders. NVIDIA strategic partnership and Omniverse integration. Facility-specific reinforcement learning creates compounding accuracy advantage. $15M ARR growing from $8.8M in 2023. Physical deployment moat cannot be replicated without equivalent operational time. | Honeywell, Siemens, Schneider Electric, and Johnson Controls are embedding AI in existing BMS platforms with massive installed base advantages. Market adoption requires physical deployment teams and customer change management. |
| Opportunities | Threats |
| Data centers projected to consume 8% of global electricity by 2030. Every hyperscale operator faces regulatory and commercial pressure to improve efficiency at existing sites before building new ones. Asia-Pacific adding over 5GW of new AI-ready capacity by 2028. | Industrial automation incumbents can bundle competing capabilities with existing long-term service contracts. AI explainability requirements in industrial control could slow deployment. |
Regulatory and compliance hurdles: Industrial control system AI must comply with IEC 62443 for industrial cybersecurity and ISO 50001 for energy management. Pharmaceutical manufacturing deployments must comply with FDA 21 CFR Part 11 for electronic records and potentially EU Annex 11 for GMP validation. Physical safety guardrails and operator override architecture are designed to address regulatory risk in advance.
Go-to-market: Direct enterprise sales to data center owners, pharmaceutical manufacturers, and industrial facility managers. NVIDIA partnership creates a channel into AI factory design engagements that embed Phaidra before a facility is operational. International expansion anchored by Singapore deployment, with Asia-Pacific as the primary growth geography given grid congestion costs and regulatory carbon pressure.
GP Lens: By 2030, AI compute infrastructure will consume a material fraction of global electricity. The companies that reduce that consumption at existing facilities — without building new hardware — are providing a service whose value compounds as compute demand grows. Phaidra’s reinforcement learning approach, built by the team that solved this problem at DeepMind and Google, produces efficiency gains that improve automatically with operational data. The NVIDIA partnership is not just capital validation. It is a distribution channel into every new AI factory being designed. At $110 million raised and $15 million ARR, this is still early relative to the size of the problem it is solving.
Pano AI — San Francisco, CA | Founded 2020
What they do: Pano AI deploys networks of ultra-high-definition panoramic cameras on ridgelines, towers, and mountaintops in fire-prone regions, integrating the camera feeds with satellite data and weather inputs through computer vision models trained on nearly 100,000 fire events. The platform detects smoke within minutes of ignition and delivers precise triangulated GPS coordinates to first responders, utilities, and emergency managers — continuous monitoring rather than the 20-to-30-minute revisit cycles of satellite-based detection. CEO Sonia Kastner and co-founder Arvind Satyam have deployed across 30 million acres in 16 U.S. states and provinces, Canada, and Australia, working with 250-plus public safety agencies and 15 major electric utilities.
Revenue: Over $100 million in customer contracts signed as of June 2025. Approximately $50,000 per station per year, covering hardware, software, maintenance, monitoring, and alerts. Portland General Electric, Arizona Public Service, Xcel Energy, and Risk Strategies are anchor customers. Two of the world’s largest insurance companies — Liberty Mutual and Tokio Marine — are both investors and strategic distribution partners.
Latest raise + backers: $44 million Series B in June 2025, led by Giant Ventures, with Liberty Mutual Strategic Ventures, Tokio Marine Future Fund, Salesforce Ventures, Initialized Capital, Congruent Ventures, and Valor Equity Partners participating. Total raised: $89 million. Giant Ventures is a climate-focused European VC. Liberty Mutual and Tokio Marine are insurers investing strategically, not Tier 1 technology VCs. None of these backers are Andreessen Horowitz or Sequoia.
IP: Multiple patents on camera-based wildfire smoke detection and triangulation, including proprietary methods for automated smoke detection, multi-camera triangulation to generate precise GPS coordinates, and the integration of AI-based smoke detection with satellite feeds and weather data. The dataset of smoke imagery, fire progression data, and alerting outcomes across nearly 100,000 fire events represents a training corpus that no competing system can quickly replicate. Physical camera infrastructure embedded in utility emergency operations creates defensibility that IP filings alone cannot capture.
Top customers: Arizona Public Service. Portland General Electric. Xcel Energy. Risk Strategies (one of the largest U.S. insurance brokers). 250-plus public safety agencies including state-level fire agencies in Oregon, California, Washington, and Colorado.
Moat and defensibility: The Bear Creek Fire in June 2024 makes the moat case concretely. Pano was the only system to detect the smoke. It provided triangulated coordinates in minutes. The helitack response contained the fire at three acres and protected a watershed serving one million people. That outcome transforms a product into emergency infrastructure. The moment a utility restructures its incident response protocol around Pano’s alert timing, the cost of switching is measured in protocol rewrites, training cycles, and years of re-calibrating detection baselines. Every fire season adds to the training dataset, widening the performance gap with any system that has not been operational for the same period.
SWOT:
| Strengths | Weaknesses |
| $100M+ in contracted revenue on $89M raised. Physical camera network creates irreplaceable data collection infrastructure. Two major global insurers as strategic investors aligns commercial incentives with risk management. 30 million acres monitored across three countries. Bear Creek Fire containment is a documented operational proof point. | Hardware-dependent subscription model has different capital efficiency characteristics than pure software. Government procurement timelines are long. Satellite competitors are advancing on revisit frequency. |
| Opportunities | Threats |
| Nearly one billion acres burned globally in 2024. Insurance industry is building wildfire risk scoring into long-term underwriting models, creating structural adoption incentive. Climate change expanding fire risk zones eastward and internationally. | Satellite-based detection programs with improving revisit intervals could reduce Pano’s relative advantage in remote areas. Extended periods of low fire activity could soften urgency for new utility adoption. |
Regulatory and compliance hurdles: Utility purchases in regulated states may require approval from state Public Utility Commissions as capital or operating expenditure. Federal fire agencies operate under GSA procurement rules. International deployments in Australia require compliance with ACMA wireless regulations. Camera installations in ecologically sensitive areas may require environmental impact reviews.
Go-to-market: Pano’s primary channels are direct enterprise sales to electric utilities and government fire agencies, supplemented by insurance broker partnerships. The Liberty Mutual and Tokio Marine investments are not just capital — they are distribution relationships. An insurer embedding Pano’s detection data into wildfire risk scoring creates a financial incentive for policyholders to deploy the technology. Geographic expansion into Europe, South America, and additional climate-vulnerable regions is targeted for 2026.
GP Lens: The wildfire detection market is an AI market with a physical infrastructure layer and a data moat that compounds over time. Pano AI has built that infrastructure in exactly the way my investment thesis demands: embedded into the operational workflows of utilities and fire agencies who now depend on the alerts for their own performance metrics. The moment a utility restructures its emergency response protocol around a Pano alert is the moment that relationship becomes effectively permanent. With $100 million in contracts, 30 million acres monitored, and insurance industry capital as both funding and distribution, Pano AI is the clearest example in this article of what I mean by irreplaceable indoctrination into infrastructure. It is not optional. It is load-bearing.
The companies featured across this series represent my personal watchlist and research interest. I have not personally invested in any of them, and nothing written here constitutes investment advice. The views expressed are my own.
