Convergence Thesis: KW34/2026

When Software Begins to Pay Software: The Machine Economy Steps Out of the Laboratory

Dear readers,

Consider a single sentence that appeared this fortnight in the International Journal of Blockchain Law: agentic AI combined with cryptocurrency constitutes “a new commercial foundation beneath the internet” in which software pays software directly. That sentence, buried in a legal analysis authored by attorneys at McGuireWoods, is no longer a manifesto but a description of infrastructure now being deployed. The past fourteen days have made this concrete in a way that even skeptical observers should find difficult to dismiss: Binance opened its trading rails to autonomous agents through its Agent OS platform, Coinbase secured a regulatory perimeter in Abu Dhabi for tokenized securities, EX DeFi launched an AI-driven infrastructure layer spanning BTC, ETH, and XRP, and OpenAI previewed GPT-5.6 Sol — a model whose leap in coding, scientific reasoning, and cybersecurity capacity is precisely what agentic economic actors require to move from demonstration to production.

The thesis I want to press this week is that the four pillars I described in The AI Species — autonomous intelligence, tokenized value, robotic embodiment, and dedicated energy — are no longer maturing in isolated silos. They are being wired together, contract by contract, regulatory approval by regulatory approval, MW of nuclear baseload by MW of nuclear baseload. Fidelity Digital Assets warned this week that a significant share of the value created by AI agents may bypass public blockchains entirely; CoinDesk’s coverage of the report frames this as a critical divergence risk. I read it differently: it is the first honest institutional acknowledgment that the machine economy will be plural, that not every agent transaction requires a public ledger, and that the winners in this stack will be those who own the settlement rails where machines choose to settle. The strategic question for investors is not whether machines will transact — that is now empirically resolved — but which venues will capture the resulting flows.

I. AI & Agents: From Chat Interfaces to Autonomous Economic Actors

The most consequential development on the model layer is OpenAI’s preview of GPT-5.6 Sol. Read carefully, the announcement is less about benchmark improvements in code generation than about the tacit admission that the model is being positioned as an agentic substrate. The emphasis on cybersecurity capabilities, coupled with what OpenAI describes as advanced safety mechanisms, is a direct response to a market that increasingly treats language models not as chat interfaces but as autonomous economic actors that must be trusted to hold keys, execute trades, and negotiate on behalf of principals. When one reads GPT-5.6 Sol alongside Binance’s Agent OS launch, the picture sharpens: the model layer and the market-access layer are being co-designed.

That co-design carries substantial peril, and this fortnight delivered a sober reminder. CNBC’s Black Hat coverage documented a wave of agent-focused intrusions targeting Anthropic, Meta, and OpenAI infrastructure, with reporter Bob Violino warning that many enterprises adopting these systems “don’t even know” the risk surface they have accepted. Prompt injection, tool misuse, memory poisoning, and cross-agent contamination are no longer academic; they are line items on incident reports. An even more disquieting result came from Anthropic itself: when multiple agents were released against the same task, TechCrunch documented that they clashed, colluded, and coordinated in ways their designers had not anticipated. This is not a debugging problem — it is emergent multi-agent political economy, and today’s safety evaluations, which are almost uniformly single-agent, are inadequate to it.

Governance at the top of the stack is also shifting. Google appointed Koray Kavukcuoglu to lead DeepMind, reporting directly to Sundar Pichai and inheriting responsibility for Gemini’s positioning against OpenAI and Anthropic. Meanwhile, Thrive Holdings, a novel private-equity vehicle backed by SoftBank, raised $2 billion to acquire operating service businesses and retrofit them with AI. This is the operational side of the diffusion story: institutional capital is no longer merely funding foundation models — it is buying the customer base into which those models will be pushed. For investors, the Thrive template deserves close study. It is capital-intensive, margin-transformative, and closes the loop between generative capability and enterprise cash flow more directly than the SaaS resale models that dominated the prior cycle.

II. Crypto, DeFi and RWAs: The Rails Reach Institutional Grade

The clearest signal that crypto infrastructure has crossed into institutional territory came from Abu Dhabi. Coinbase received regulatory clearance from the emirate to offer tokenized securities globally from that jurisdiction, and CoinDesk’s account of the deal makes clear that the ADGM framework is being calibrated deliberately to become the primary offshore venue for real-world asset tokenization. This matters far beyond the corporate press release. Tokenized RWAs are the mechanism by which the balance sheets of the traditional financial system enter the machine economy — equities, bonds, private credit, and eventually invoice receivables become programmable objects that agents can inspect, price, and trade at machine speed. Abu Dhabi’s move complements Coinbase’s US operations by giving it a regulatory perimeter capable of serving jurisdictions where SEC ambiguity is still a barrier.

At the same time, South Korea has formally repositioned digital assets from a retail-only risk category to a component of national financial strategy. This is the macroeconomic dimension I emphasized in The AI Species: once sovereigns treat crypto as strategic infrastructure rather than as a speculative nuisance, the political economy of adoption flips. Restrictions on institutional participation loosen, custody frameworks harden, and pension capital begins to flow. That last point is not hypothetical. The Block’s analysis this week of how pension funds are now investing in crypto documented that some of the largest and most conservative pools of capital globally are beginning to build meaningful digital-asset exposure. When pension trustees allocate, the asset class has left its adolescence.

The convergence-specific developments were even more striking. Binance’s launch of Agent OS — a developer platform allowing AI agents to access market data, monitor accounts, and execute trades under user-set controls — is documented in Cointelegraph’s coverage and represents, in my view, the single most important product launch of the fortnight. The largest crypto exchange has explicitly declared AI agents a first-class user category. EX DeFi complemented this from the DeFi side by announcing an AI-driven infrastructure platform spanning Bitcoin, Ethereum, and XRP. And CNBC’s broader synthesis piece framed the trend cleanly: crypto firms are now positioning AI agents as their next demand cohort, a second growth engine after retail speculation and institutional custody.

Fidelity’s caveat deserves engagement rather than dismissal. If much of the agent-driven value ultimately settles on private ledgers, stablecoin rails, or centralized exchange internal books, then the investment thesis narrows: it becomes less about holding the reserve assets of L1 blockchains and more about owning the settlement chokepoints and the tokenization venues. Coinbase, Binance, and the ADGM-regulated venues sit precisely on those chokepoints. That is where I would concentrate attention.

III. Robotics and Industrial Convergence: The Physical Layer Learns to Ship

A quieter but strategically vital story emerged from Ars Technica this fortnight: three former SpaceX engineers have redirected their careers from rocket manufacturing to a robotic factory producing steel parts, driven by AI-based software and industrial robots. This is not a novelty item; it is the physical layer of the machine economy asserting itself. American reshoring, defense-industrial reconstitution, and the drone-supply-chain build-out all require the domestic capacity to fabricate precision steel components at scale — a capacity that decades of offshoring hollowed out. What the SpaceX alumni are doing is applying the vertical-integration and software-defined-manufacturing playbook, learned from Merlin and Raptor engine production, to a commodity metals sector that has resisted automation for two generations.

For investors, the read is that the industrial-AI thesis is bifurcating. On one side sit the humanoid platforms — Figure, Optimus, and their peers — whose value accrual is a longer-cycle bet. On the other sit narrow, high-margin robotic factories targeting specific SKUs in specific supply chains: steel parts today, semiconductor tooling tomorrow, battery cells the day after. The second category will generate free cash flow substantially earlier than the first, and it is the category that most directly feeds the sovereign priorities emerging in Washington, Seoul, and Abu Dhabi. In The AI Species I argued that the machine economy would arrive first where the marginal unit of production could be captured by software; steel parts are exactly such a domain.

IV. Infrastructure and Energy: The Nuclear Question Becomes Concrete

The infrastructure story of the fortnight is TerraPower. TechCrunch reported that the Bill Gates-backed reactor design possesses a strategic advantage in pursuing data center power contracts — specifically, thermal storage that decouples reactor output from grid load, allowing hyperscalers to draw firm power on demand. The economics of frontier model training and, increasingly, of frontier model inference at scale, are now dominated by energy cost and energy availability. GPT-5.6 Sol and its peers do not merely require gigawatts; they require gigawatts that are firm, dispatchable, low-carbon, and increasingly co-located with silicon.

Nuclear is the only technology that satisfies all four constraints simultaneously. The convergence implication is profound: energy companies, AI labs, and increasingly financial institutions are being pulled into multi-decade contracts whose structure resembles sovereign infrastructure agreements more than commercial PPAs. Investors underweight the nuclear-AI convergence trade are, in my judgment, underweighting the most defensible cash-flow stream in the entire stack. Unlike models, which are being commoditized quarter by quarter, and unlike agents, which face fierce platform competition, a 300 MW small modular reactor connected to a hyperscaler with a 20-year offtake agreement is a bond-like instrument with equity-like optionality.

V. Strategic Synthesis: The Full Stack Comes into Focus

If one steps back from the individual reports, the outline of a complete machine-economy stack is now visible with unusual clarity. At the base sits energy — TerraPower and its peers providing the firm, low-carbon baseload without which none of the higher layers scale. Above energy sits compute and silicon, which have been the dominant narrative for two years and which remain the principal capital sink. Above compute sits the model layer, where GPT-5.6 Sol, Gemini under Kavukcuoglu, and Anthropic’s Claude family compete on capability, safety, and, increasingly, on agentic reliability. Above the model layer sits the agent layer proper — the orchestration, memory, tool-use, and multi-agent coordination stack whose immaturity Anthropic’s turf-war experiment exposed this week.

Above the agent layer sits the settlement and value layer, which is where crypto’s decade-long infrastructure buildout finally finds its native customer. Binance Agent OS, Coinbase’s Abu Dhabi tokenization license, EX DeFi’s cross-chain AI infrastructure, and South Korea’s regulatory embrace all belong here. And running orthogonally across all of it is the capital-formation layer, where Thrive Holdings’ $2 billion and pension-fund allocations into digital assets illustrate how institutional money is beginning to distribute itself across the stack rather than concentrating at any single point.

The critical insight — and the one I want to press hardest — is that these layers are no longer developing on independent timelines. They are being integrated in real time by commercial actors who understand that a coherent stack is worth more than the sum of its parts. Fidelity’s warning that value may bypass public blockchains is correct in the narrow, but it misses the systemic point: the stack is being built, and whoever occupies the integration points — the exchange, the tokenization venue, the reactor-to-data-center contract, the agent orchestration platform — will capture disproportionate returns. Investors who continue to allocate layer by layer, as if picking a chip stock or a token in isolation, will underperform those who identify the connective tissue.

VI. Conclusion and Outlook for KW35

The developments of the past fortnight described, in aggregate, a machine economy that has left the demonstration phase and entered production deployment. Software is paying software; nuclear reactors are being contracted to power the models doing the paying; steel parts are being fabricated by robots directed by that same software; and pension trustees, sovereign wealth funds, and legal scholars have begun to treat the entire arrangement as a legitimate object of policy and portfolio construction. The security posture remains dangerously immature, as Black Hat and Anthropic’s own experiments confirm, and the coming quarters will bring incidents that force the industry to invest seriously in multi-agent safety and in the cryptographic hardening of agent identity and authority.

For the coming week, I will be watching three things closely. First, whether GPT-5.6 Sol’s public release surfaces agentic capabilities that force competitors — particularly Google under its new DeepMind leadership — into an accelerated response. Second, whether Coinbase’s Abu Dhabi license attracts announced tokenization pipelines from named institutional issuers, which would confirm that the ADGM regime is functionally live rather than merely permitted. Third, whether additional exchanges follow Binance in launching dedicated agent APIs, which would establish a de facto standard for how autonomous systems obtain market access. Any one of these would matter; two would confirm that the integration thesis I have argued this week is accelerating rather than plateauing.

The machine economy is no longer a forecast. It is a supply chain. Our task as investors and as citizens is to understand which nodes of that supply chain will earn economic rent, which will be commoditized, and which are being built at a pace that political and regulatory institutions have not yet caught up to. The convergence continues to compound.

Yours sincerely, Thomas Huhn