Connectivity Is the New Gold

1. Executive summary

Thesis. Connectivity is becoming the invisible growth constraint of the Cloud and AI economies, and the highest-asymmetry way to own it is picks and shovels: the connection, coordination, verification, settlement, and deployment layers of a buildout that centralised balance sheets cannot complete alone. The more agents (humans, physical agents, AI agents) are created, the more strain they put on existing infrastructure. Blockchain coordination is the only mechanism that can organise hardware and software into enterprise-grade infrastructure at global scale. Stablecoins showed how crypto wins a financial layer. It disappears into the product and customers may already not know that they are using blockchain. I believe the physical layer of the internet is next.

Multi-sided market problem:

1. Supply: a twenty-billion-device infrastructure globally deployed to humans, roughly 60% of it consumer-owned, is the one infrastructure pool hyperscalers structurally cannot recruit, and its distribution already perfectly matches the consumer [1][2][3].
2. Demand: physical AI and agentic traffic break the agent-data-center-agent architecture on bandwidth and latency alone, forcing inference and settlement to the edge [4]-[7].
3. Value accrual: in every infrastructure cycle the raw resource commoditises and margin concentrates in contracted service and coordination on top; crypto is running the same script, and the sector’s own 2025 data confirms the turn [8][9].

Four pillars where I believe in the biggest pay-offs:

1. Consumer-device coordination (crowdsourcing storage, compute and connectivity hardware).
2. Expansion of the existing infrastructure layer (new solutions targeting connectivity bottlenecks).
3. Physical-AI infrastructure (verifiable digitalisation of physical world data, connectivity between physical devices and data center, legal solutions to enforce digital decision in physical world).
4. Machine, Agent and Edge Infrastructure payment rails.

2. Market definition and sizing

This addresses the infrastructure layer of internet services, including compute, bandwidth, data access, storage, and the payment rails between them at the point where it meets AI-era demand.

3. Why now?

The AI and Cloud buildout is guided by CapEx. The four largest hyperscalers plan close to $700B in combined 2026 capital spending on CNBC’s count [11]; Goldman Sachs aggregates the same four at ~$725B, up 77% from a record ~$410B in 2025, and projects $5.3 trillion for FY2025-2030 with a $7.6 trillion baseline for 2026-2031 across compute, data centers, and power [12]. Oracle adds roughly $50B, and the Stargate programme targets $500B on its own [16]. Underneath it all, worldwide public cloud spending reaches $723.4B in 2025, growing 21.5% in a single year on Gartner’s forecast [10].

I see the following inflection points as the markers for the opportunity:

1. The constraint has gone physical. Roughly 40% of announced AI data center projects face construction delays from power infrastructure, not GPU supply [17].
2. Spend is centralising while workloads decentralise. Training concentrates; inference, data collection, and agent traffic disperse to wherever vehicles, robots, devices, and users are. Gartner’s projection that 75% of enterprise-generated data would be created and processed outside traditional data centers by 2025, up from about 10% in 2018 [18], set the direction; the $378B edge spending forecast confirms the migration is funded [13].
3. The regulatory clarity is finally coming through. The GENIUS Act gave stablecoin settlement a US federal framework [19], with Clarity and other regulations coming in the space, more and more crypto rails would be mainstream adopted.

At the same time, direct disruption of industries with this CapEx intensity is close to impossible; nobody outbuilds a $200B annual budget. The asymmetric entry is the existing infrastructure coordination: capacity that already exists in consumers’ hands, and coordination markets their business model cannot operate.

4. The pattern that crypto adoption is about to repeat

Value accrual follows a script seen in Cloud and Mobile before. Raw capacity commoditises; margin concentrates in whoever sells guaranteed, contracted service and coordination on top. AWS closed FY2025 at 18% of Amazon’s revenue and 57% of its operating income [20]; US tower operators quadrupled revenue while per-user prices fell 30% and traffic rose 341x, and T-Mobile paid up to $1.35B for an MVNO that owns no towers at all [20]. Crypto infrastructure is approaching the same script, and 2025 was the year its data confirmed it. Leading DePIN revenues decoupled from token prices. Helium’s onchain revenue grew roughly 8x while its token fell 77% [21] while the speculative 2018-2022 cohort sits 94-99% below highs [8], and Helium Mobile being sold without reaching profits. Storj’s Chapter 11, restructuring like a normal service business to keep serving customers, and so on [20]. Major Crypto and DePIN players who succeeded in customer acquisition are moving away from token narratives, while the biggest winners are the ones who built services on top, e.g. Hyperliquid, Kalshi, PYUSD.

The supply side is where blockchain is necessary; the demand side is where it doesn’t matter or even makes things worse. Enterprises will keep buying compute, data, and bandwidth regardless of the rails underneath. I watch this weekly, because major AI companies buy from my own network for service quality at 60-75% below traditional cloud cost, not for decentralisation [22]. The moment no customer talks about the blockchain, the product has succeeded. Stablecoins crossed that line for money: $46T in total volume (~$9T adjusted) in the twelve months to October 2025, up 106% [19], and ~$10.2T trailing adjusted volume by mid-2026 [14] with issuers among the largest holders of US Treasuries and users who increasingly never notice the crypto. The physical layer crosses the same line through the four pillars below.

5. Investment pillars

5.1 Consumer-device coordination

Thesis. Consumer devices are the one infrastructure pool hyperscalers cannot recruit, and token coordination is an effective crowdsourcing mechanism.

Evidence. Connected IoT devices reached 18.5 billion at end-2024, with 2025 estimates of 19.8-21.1 billion depending on tracker methodology, heading toward 39-41 billion by the early 2030s [1][3]; mobile phones add ~8.6 billion connections counted separately [2]; roughly 60% of the device base is consumer [3]. This fleet is already deployed, powered, housed, and connected in perfect service locations: in homes, last-mile networks, and residential IP space across 130+ countries that no data center can occupy by definition. Recruiting it is not a sales problem but a mechanism-design problem: enrol, verify, meter, and pay millions of anonymous owners with no trust relationship. Token coordination can solve it; hyperscalers cannot.

Entry criteria. Supply categories with verified enterprise demand (bandwidth, residential data access, distributed inference, sensor coverage); verification mechanisms strong enough to underwrite an SLA; subsidy ratios on a credible sunset path.

Anti-patterns. No long-term enterprise customer adoption. Inability to run SLAs, or compete with traditional cloud on service quality.

5.2 Expansion of the existing infrastructure layer

Thesis. New solutions targeting existing connectivity bottlenecks will appear as the market and legal frameworks mature. Stablecoins disrupted payment rails, infrastructure should be next.

Evidence. The edge is millions of heterogeneous locations owned by nobody in particular, absorbing $378B of forecast spend by 2028 [13] as the 75%-outside-the-data-center data shift plays out [18]. BitTorrent demonstrated crowdsourced capacity allocation without a single source of truth two decades ago; the infrastructure layer by itself would be distributed via new technical solutions that are closer to edge like never before. Tools for coordination already created: stablecoin settlement (machine-speed, sub-cent, jurisdiction-agnostic, now GENIUS-legal [19]) and onchain coordination (permissionless supply enrolment, verifiable metering, programmatic revenue share).

Entry criteria. Unique technical or network coordination solution that is not implementable by existing or traditional infrastructure. Clever usage of settlement, metering, verification, and orchestration layers for connecting fragmented supply and enterprise demand.

Anti-patterns. Economics tied to token success. Overcompensation of infrastructure buildout. Lack of clear enterprise adoption. Lack of clarity on who will be paying for that solution.

5.3 Physical-AI infrastructure

Thesis. Further growth of agentic and physical AI requires further breakthroughs in verifiable digitalisation of physical world data, connectivity between physical devices and data center, legal solutions to enforce digital decision in physical world.

Evidence. Sensor pipelines on a single autonomous vehicle generate up to 4TB per hour, rising toward 19TB per hour at higher autonomy on McKinsey’s estimates, with 3-40 Gbit/s onboard processing [4]. In production, Waymo retains ~25TB of logged data per vehicle per day [5], which can overwhelm any city network if scaled. An automated factory with sensor-dense systems generates 4-7TB per hour [6]. Streaming raw sensor data from millions of production vehicles “would crash existing cellular networks” [7]. Cloud round trips of 800-2,400ms cannot serve control loops needing sub-10ms responses [6]. Inference therefore has to move to local and regional infrastructure, while DePIN already has the playbook for deploying globally distributed, individually owned, economically incentivised node fleets. Hopefully, a decade of proof-of-work and DePIN operations produced the field knowledge that is relevant. Hardware distribution, uptime incentives, verification of physical work, operator payback economics are the existing tools that the edge buildout would need at industrial scale.

Entry criteria. Edge and regional inference networks. Bandwidth and connectivity layers. Node-deployment and fleet-management tooling building out AI equipment. Physical-world data pipelines (driving, robotics, sensors) with clean rights chains. Enforceable SLAs and other legal solutions that enforce the connection between physical and digital world.

Anti-patterns. Promise of outperforming centralised hyperscalers at compute. Capacity aggregation with no adoption clarity. Lack of non-Web3 use-cases.

5.4 Machine, Agent and Edge Infrastructure payment rails

Thesis. Machine-to-machine commerce might adopt crypto rails first. The value concentrates in the settlement and coordination layers that connect the money and the resource via execution pipeline.

Evidence. x402 (HTTP 402 plus stablecoins) processed over 169 million payments across ~590,000 buyers and ~100,000 sellers in its first year [15]. Governance moved to a Linux Foundation home in April 2026 with 22 members including Visa, Mastercard, American Express, AWS, Google, Microsoft, Anthropic, Stripe, Circle, and Shopify [23]. Cloudflare wired a deferred-payment scheme into pay-per-crawl [24]. Cloudflare reports that 52% of crawler requests are now AI-related, up from 22% in spring 2025 [15]. The internet’s traffic mix is flipping from humans to machines. Agents will soon be paying for compute, data, and bandwidth and would need rails for that. The silent winners would run that layer.

Entry criteria. Agent-native infrastructure, wallets and spend-policy automations. New ways of metering and verification for machine-bought resources. Solution succeeds exactly when end customers do not know what they are using: coordination solved first, then nobody talking about the blockchain underneath.

Anti-patterns. Consumer-facing “pay with crypto”.

6. Who will win?

To validate projects, I would love to see the following metrics:

1. Demand-side revenue, externally verifiable. Fiat or stablecoin revenue from customers who are not token holders. Ideally, onchain-provable or internally audited, never blog-post-reported [8][9].
2. Incentive spend per revenue dollar, trending down. Network incentives if any should be comparable to demand revenue, with payouts that survive a flat token price. The ideal case is settlement in stablecoins, which does not depend on token price fluctuations.
3. Enterprise PMF, not community PMF. PoC-to-paying enterprise conversion, recurring revenue share, SLA capability, long-term contracts with minimum capacities [9].
4. Supply-side necessity. Network must solve a coordination problem that hiring cannot. If the supply could be procured conventionally, the network is not defensible at large valuations or major competition.
5. Ability to disappear for end consumer. Does end-consumer need to know anything about your technical stack or can just use product/service straight away?

7. Risks and falsifiers

What would make this thesis wrong:

1. Mainstream absorption. AWS shipped x402 support in CloudFront at general availability while Cloudflare’s gateway sat waitlisted [15]. Rails can be adopted by existing edges faster than new networks form around them. Mitigant: the supply-side coordination problem (Pillar 1) remains crypto-native even if payment rails commoditise. Falsifier: machine-payment volume concentrating durably inside incumbent service providers.
2. Subsidy reflexivity. Emission-financed supply collapses when the token subsidy sunsets before organic demand arrives. Watch: sector-wide incentive-spend and CAC ratios, even Monday.com and similar SaaS were poisoned with the same dynamic.
3. On-device efficiency outruns the edge network. Model compression is advancing fast enough that some inference may stay on the device itself, shrinking the intermediate edge layer this thesis expects to grow. Watch: the split of inference workloads between device, edge, and region in production robotics and autonomous driving deployments.
4. Demand concentration and geography. Much of today’s DePIN demand clusters in a small set of AI buyers, heavily US and Asia. A revenue base with concentration and geopolitical exposure I know from the inside. Watch: legal grounds, startup and founder origins together with recent legal activities across the sector.
5. Regulatory frameworks. Residential data access and scraping operate in legal grey zones; consumer-device consent standards could reprice the supply side, and CLARITY-era reporting discipline is a must.

8. Investment opportunity

If you are building in this space and align with the thesis, I would be happy to consider an angel investment. Please share with me your blurb and deck, and I will reach out if I would like a follow-up conversation.

References

[1] IoT Analytics, “Number of connected IoT devices” (ongoing series; May 2026 update). 18.5B connected IoT devices end-2024; 21.1B forecast end-2025; ~39B by 2030. https://iot-analytics.com/number-connected-iot-devices/

[2] Statista/DataReportal, connected devices worldwide by type (Oct 2025). Mobile phones ~8.62B connections, tracked separately from IoT counts. https://www.statista.com/statistics/1559435/connected-devices-worldwide

[3] Statista, IoT connections worldwide 2022-2034. 19.8B (2025) growing to 40.6B (2034); consumer segment ~60% of devices. Note: [1] and [3] differ on 2025 (21.1B vs 19.8B) by tracker methodology; the text quotes the range. https://www.statista.com/statistics/1183457/iot-connected-devices-worldwide/

[4] Edge AI and Vision Alliance (Mar 2025), citing McKinsey, “Rethinking Car Software and Electronics Architecture”. A single AV generates up to 4TB/hour, toward 19TB/hour at higher autonomy; 3-40 Gbit/s onboard pipelines. https://www.edge-ai-vision.com/2025/03/optimizing-edge-ai-for-effective-real-time-decision-making-in-robotics/

[5] Introl, “Autonomous Vehicle AI Infrastructure” (2026). Waymo ~25TB of logged data retained per vehicle per day; sub-10ms safety loops. Generation [4] and retention [5] are different measures; the text cites them as such. https://introl.com/blog/autonomous-vehicle-ai-infrastructure-edge-cloud

[6] Moduledge, “Edge Data Centers for Robots” (Jun 2026). 4-7TB/hour for sensor-dense automated systems; cloud round trips 800-2,400ms vs sub-10ms control loops. https://www.moduledge.com/blog/edge-data-center-robots

[7] DataM Intelligence, “Edge Computing in Autonomous Vehicles: Level 4 Architectures” (Jun 2026). Streaming raw sensor data from millions of production vehicles “would crash existing cellular networks.” https://www.datamintelligence.com/blogs/edge-computing-in-autonomous-vehicles-level-4-architectures

[8] Messari and Escape Velocity Ventures, “State of DePIN 2025” (Jan 2026). ~$10B circulating market cap; ~$72M FY2025 onchain revenue; leaders at 10-25x revenue vs 1,000x+ in 2021; 2018-2022 cohort down 94-99% from highs. https://messari.io/report/state-of-depin-2025

[9] K. Tkachuk, “DePIN Is Dead. Long Live DePIN.” (4 Apr 2026). Part 1 of the sector series: only two pre-2023 projects crossed $1M annualised non-token revenue; supply-subsidy failure data (Helium, Filecoin, Livepeer, Akash); the survivors’ shift to fiat-billed, SLA-backed services; “the service layer always eats the value.”

[10] Gartner press release (19 Nov 2024). Worldwide public cloud end-user spending $723.4B forecast for 2025, up 21.5% from $595.7B in 2024. https://www.gartner.com/en/newsroom/press-releases/2024-11-19-gartner-forecasts-worldwide-public-cloud-end-user-spending-to-total-723-billion-dollars-in-2025

[11] CNBC, “Tech AI spending approaches $700 billion in 2026” (6 Feb 2026). Four hyperscalers ~$700B combined 2026 capex. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html

[12] B. Sozzi, Yahoo Finance, reporting Goldman Sachs (A. Lynam) aggregation (3 Jun 2026). $725B combined 2026 capex for the four hyperscalers, up 77% from $410B; $5.3T FY2025-2030; $7.6T baseline 2026-2031 across compute, data centers, and power. Consistent with [11]; same four companies, later estimate date. https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html

[13] Forbes Technology Council (12 Dec 2024), citing IDC. Global edge computing spend forecast $378B by 2028. https://www.forbes.com/councils/forbestechcouncil/2024/12/12/2025-it-infrastructure-trends-the-edge-computing-hci-and-ai-boom/

[14] Visa Onchain Analytics Dashboard (with Allium). Live dashboard for stablecoin supply and adjusted volume. The ~$10.2T trailing-12-month adjusted volume and ~$322B supply readings (Jul 2026) as captured in Solana Compass reporting. [19] and [14] are different vintages of the same series: ~$9T (Oct 2025) growing to ~$10.2T (mid-2026). https://visaonchainanalytics.com/ and https://solanacompass.com/news/visa-onchain-analytics-reports-record-179-trillion-in-adjusted-stablecoin-volume-for-june-2026

[15] InfoQ, “Cloudflare and AWS Embed x402 Agent Payments at the Edge” (Jul 2026). x402: 169M payments across ~590k buyers and ~100k sellers in year one; 52% of crawler requests AI-related (Jun 2026) vs 22% (spring 2025); AWS CloudFront x402 at general availability vs Cloudflare gateway waitlist. https://www.infoq.com/news/2026/07/cloudflare-aws-x402-micropayment/

[16] Futurum Group, “AI Capex 2026: The $690B Infrastructure Sprint”. Oracle ~$50B 2026 capex; Stargate $500B programme. Its five-company aggregate reflects earlier-quarter guidance; the four-company series in [11][12] is primary here. https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/

[17] NextWaves Insight (Jun 2026). ~40% of announced AI data center projects delayed by power-infrastructure bottlenecks. Secondary source; treat directionally. https://nextwavesinsight.com/hyperscaler-ai-capex-microsoft-google-amazon-meta-2026/

[18] Gartner, “What Edge Computing Means for Infrastructure and Operations Leaders” (S. Rao). The projection that 75% of enterprise-generated data will be created and processed outside traditional data centers/cloud by 2025, up from ~10%. https://www.gartner.com/smarterwithgartner/what-edge-computing-means-for-infrastructure-and-operations-leaders

[19] a16z crypto, “State of Crypto 2025” (Oct 2025). Stablecoins $46T total / ~$9T adjusted volume in the 12 months to Oct 2025, +106% YoY; issuers holding >$150B of US Treasuries; >1% of US dollars tokenised; GENIUS Act framing. https://a16zcrypto.com/posts/article/state-of-crypto-report-2025/

[20] K. Tkachuk, “DePIN is Dead. Long Live Infrastructure-as-a-Service.” (19 Aug 2026). Part 2 of the sector series, with primary sources for: Amazon Q4 2025 earnings release (5 Feb 2026), FY2025 AWS revenue $128.7B (18% of Amazon revenue) and operating income $45.6B (57% of Amazon’s $80.0B total); US tower operator revenue and CTIA ARPU and traffic data; T-Mobile’s acquisition of Mint Mobile’s parent; the Helium Mobile sale; Storj’s Chapter 11.

[21] Cointelegraph, “DePIN revenues grow despite token losses” (Jan 2026). Helium onchain revenue ~8x in 2025 while HNT fell 77%; GEODNET revenue 1.7x while GEOD fell 41%; DePIN Leaders Index thresholds ($500k+ ARR, $30M+ raised). https://cointelegraph.com/news/depin-revenues-grow-despite-token-losses

[22] Titan Network: whitepaper and operating metrics. https://whitepaper.titannet.io/en

[23] Eco, “x402 Protocol Explained” (2026). Linux Foundation formalisation 2 Apr 2026; 22 launch members incl. Visa, Mastercard, American Express, AWS, Google, Microsoft, Anthropic, Stripe, Circle, Shopify. https://eco.com/support/en/articles/14839402-x402-protocol-explained

[24] Coinbase, “Coinbase and Cloudflare Will Launch the x402 Foundation” (23 Sep 2025). Foundation announcement; Cloudflare deferred-payment scheme for pay-per-crawl. https://www.coinbase.com/blog/coinbase-and-cloudflare-will-launch-x402-foundation