Last updated on August 17th, 2026 at 11:15 pm
As we say every week in AI: what a week!
For most of the generative AI boom, intelligence itself was the scarce resource. Only a handful of companies could build genuinely capable models. Access was expensive. Every major release was an event. Benchmarks moved markets, and the industry organized itself around one central question: who has the smartest AI? That question is becoming less important.
Capable models are proliferating. Open-weight systems are improving, even overtaking. Inference costs continue to fall. Smaller models can perform work that recently required the frontier. Models can increasingly be routed, combined and connected to the same tools and data. Intelligence is becoming a commodity.
“Commodity” doesn’t mean identical. It means capable intelligence is becoming less scarce, less concentrated, more abundant, competitively priced and replaceable for a growing number of tasks. Intelligence itself becomes lower value.
As that happens, the ecosystem and value chain changes. Value migrates to the scarcity layers surrounding models: proprietary data, compute, energy, capital, distribution, orchestration, permissions, relationships, ecosystems, and trust.
This week’s developments all point in that direction.
THE MACRO: Abundant Intelligence Creates New Bottlenecks
Nvidia provided the clearest evidence from the infrastructure side.
The company announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR intended to mobilize more than $500 billion in third-party capital for AI compute infrastructure. The initiative addresses one of the industry’s most obvious bottlenecks: somebody has to finance the enormous physical system underneath AI.
Data centres require chips, land, power, cooling, networking and extraordinary amounts of capital. Nvidia CEO Jensen Huang has gone further, describing computing capacity as an emerging asset class.
That marks a significant evolution for an industry that, only a few months ago, was primarily discussing models, benchmarks and tokens. AI infrastructure is becoming something institutional investors can finance alongside energy, real estate and transportation.
The intelligence may be getting cheaper. The infrastructure required to produce and deliver it is becoming more valuable.
MACRO: Meta Reopens the Intelligence Stack
Meta approached the same transformation from the other direction. Mark Zuckerberg used the launch of Muse Glimmer to make a renewed case for open-weight AI and for putting increasingly capable intelligence directly into the hands of users.
If capable models can run locally on consumer hardware, people and businesses do not have to rent all of their intelligence from a small number of centralized providers. They can possess it, modify it, deploy it and build directly on top of it.
Meta is doing what it has to do and trying to play catch up. The companies therefore appear to be making opposite bets. Nvidia is helping finance enormous centralized compute infrastructure. Meta is trying to distribute capable intelligence outward.
Both strategies, however, depend on the same underlying assumption: AI intelligence is going to become abundant and ubiquitous.
Nvidia is betting (correctly) that demand will become large enough to support a new global infrastructure class. Meta is betting that models will become common enough to function as platforms. The strategy took a hit when its acquisition of Manus was unwound by the Chinese government, and not coincidentally? this week Manus announced that it will once again be delivering its technology to the world (with an odd data interruption). Meta is doubling down on a strategy it had largely abandoned: trying to demonstrate its utility in open weights, and the Trump administration, recognizing that America has lost the edge, in a new hands-on regulatory capacity has already shown signs of supporting this direction.
In both cases, competitive advantage shifts toward what surrounds the model: hardware, capital, distribution, proprietary data, interfaces and ecosystems.
THE MICRO: In Ecosystems, The Model Becomes a Component
A couple of B2B marketing announcements this week illustrated the same transition: increasing integration between AI and industry.
Unbounce and Getty Images added support for the Model Context Protocol, or MCP, which is emerging as a standardized way for AI systems to connect with tools, services and data. Standardized interfaces reduce the cost of connecting applications to different AI systems. They make models easier to integrate, route and replace. The technical details help web developers and marketers who want to test landing pages or source images at scale (disclosure: I have sold photography to Getty Images), and the economic implications are that this becomes faster, cheaper and easier to iterate.
The model begins to behave less like the product itself and more like one component within a larger ecosystem. Value then shifts toward the layers that determine how intelligence reaches the market: distribution, applications, developer communities, proprietary data and established customer relationships. Orchestration remains part of that layer, selecting which model performs a task, deciding what context it receives, controlling the tools it can access and setting boundaries around what it is permitted to do.
MCP is one expression of this shift, not its universal endpoint. It is one of several emerging, partly competing and partly complementary protocols designed to connect models, tools, data and other agents. MCP primarily connects agents to tools and data, while A2A, for example, governs communication between agents. The Linux Foundation explicitly describes the two as complementary. (Linux Foundation). Their importance lies less in making every model equivalent than in reducing the cost of assembling and changing the systems around them. As interoperability improves, model differences become easier to arbitrage, but distribution becomes harder to dislodge. The companies that control the dominant interfaces, integrations, workflows and developer networks may capture more durable value than the companies supplying any individual model.
Every meaningful MCP integration is therefore a developer milestone, but the larger commoditization signal is the growth of interoperable ecosystems themselves. Intelligence may become easier to source and substitute; the channels through which it reaches users, businesses and other agents may become more valuable.
AGENTIC: Intelligence Gets Hands
Interchangeable intelligence becomes much more consequential once it can act. This has recently dominated the news because of the way in which agentic systems pursue goals, but one of the most extraordinary examples we’ve seen emerged from Australia this week after a user asked an AI assistant to help book a gym class.
The OpenClaw-based agent eventually discovered weaknesses in the gym’s booking system and manipulated the system to improve its owner’s position.
The agent did not require extraordinary intelligence. It required access, an objective and insufficiently constrained authority.
As AI moves from answering questions to operating browsers, calling APIs, editing databases, making purchases and interacting with software, raw model capability becomes an increasingly incomplete measure of risk.
The most important questions are not ideological, they should be operational:
- What can the system reach?
- What is it permitted to change?
- Which actions require explicit human authorization?
- Can its actions be reconstructed and independently verified?
An agent doesn’t need to plot against humanity to cause harm. It only needs to act in conditions of sufficient ambiguity. many of the examples shown so far have had unclear instructions errors on the part of the deployment of the agent or unrestricted access to critical elements, like the Internet. A capable model with an ambiguous, poorly defined objective and access to the wrong tools is enough.
The immediate agentic risk is less exotic than many science-fiction scenarios suggest. It is the ordinary deployment of increasingly available intelligence into systems with weak permissions, inadequate monitoring and unclear accountability.
Got your tickets for MAICON 2026, October 13 to 15 in Cleveland yet?
The Marketing AI Conference returns to the Huntington Convention Center of Cleveland for its seventh year, hosted by the Marketing AI Institute. Three days on what marketing teams are actually deploying: tool evaluation and ROI measurement, content operations, personalization, governance and change management, and case studies from CMOs and operators running AI inside their organizations.
The AI for CMOs Summit runs the morning of October 13, with half-day workshops that afternoon. Main conference keynotes and breakouts are October 14 and 15. Passes come as Main Conference, All-Access with the workshop, or All-Access plus on-demand main stage recordings. Groups of four or more get 10% off, and there are custom rates above ten.
B2BNN is a media partner of MAICON 2026. https://hubs.li/Q049vD6T0 Use code AKRON150 for a discount on your registration!
THE POLICY: Governments Regulate the Layers Around Intelligence
August 2 marked another transition as Article 50 transparency requirements under the EU AI Act began applying. Providers must now ensure users know when they are interacting with certain AI systems and provide machine-readable marking for certain AI-generated or manipulated content. Deployers also face disclosure requirements involving deepfakes and some AI-generated public-interest material.
Disclosure, provenance, watermarking and AI-content detection are becoming compliance infrastructure.Europe’s focus is expanding beyond what a model can produce to how AI-generated material is identified, presented and traced.
Washington is following a very different and less coherent regulatory path, but the Trump administration’s approach is becoming increasingly difficult to characterize as simple deregulation. The administration remains committed to accelerating American AI development. It is also becoming much more willing to use federal procurement, national-security authority and industrial policy to influence how AI develops. A June national-security directive called for rapid adoption of frontier models across the national-security enterprise, the use of multiple suppliers, high-security compute facilities, stronger reliability and controllability requirements, and the creation of an AI National Security Strategic Reserve.
The emerging position is interventionist in a distinctly American way: Accelerate development. Centralize federal authority. Protect American technological leadership. Use procurement and national-security power to shape corporate behaviour.
Europe and the United States are building very different systems. Europe is emphasizing transparency and compliance. Washington is emphasizing strategic control, domestic leadership and federal power.
They increasingly agree on one underlying principle: AI has become too consequential to leave unregulated. The EU has demonstrated with GDPR that it will not hesitate to enforce its regulations. Some of the privacy violation fines have been significant and substantial. The industry is rightfully paying attention.
MARKET BUBBLE WATCH: Risk Moves Into Finance
The Inverted Bubble Holds, with Sovereignty Implications
The conventional AI-bubble theory expects speculative demand to disappear, leaving excess capacity and collapsing valuations behind it. This week’s evidence points in the opposite direction. CoreWeave reported second-quarter revenue of $2.575 billion, up 112 percent year over year, with approximately $104 billion in revenue backlog and more than $25 billion in additional customer commitments secured early in the third quarter. The company raised its expected 2026 capital expenditures to between $35 billion and $39 billion because demand continues to exceed the capacity it can bring online. Demand is not disappearing. That does not disprove the inverted-bubble thesis. It is the thesis. The inverted bubble begins with real, operational demand combined with prices that do not fully reflect the cost of providing the product.
Providers must continually finance new infrastructure, absorb the difference and borrow against future revenue to purchase hardware whose economic life may be shorter than the debt used to acquire it. Under those conditions, a company can sell everything it builds, grow extraordinarily quickly and become more financially exposed with each new increment of demand.
Source: Rohan Paul on X
CoreWeave’s results show that mechanism clearly. The company spent $9.4 billion on capital expenditures during a quarter in which it generated $2.575 billion in revenue. It recorded $640 million in net interest expense and a $626 million net loss. Its full-year capital-spending guidance is now roughly three times its expected annual revenue. Nvidia’s new $500B partnerships (these numbers are stupefying) reinforce this. The proposed platforms are intended to mobilize more than $500 billion in third-party capital for AI infrastructure. The risk has not migrated from demand into finance. Finance was always the structure supporting the demand. What is changing is who is being asked to carry the exposure as the buildout continues. None of this establishes that every proposed data centre is economically justified.
Contracted demand at one major operator is not a jurisdiction-by-jurisdiction demand model, and backlog does not prove that revenue generated farther up the stack will ultimately support the infrastructure beneath it. For now, however, the classic capacity-bubble scenario is not playing out. The immediate vulnerability is not that nobody wants AI compute. It is that enormous real demand is being served through a capital structure that may not earn adequate returns before the hardware depreciates and the financing matures.
If the inverted bubble corrects, AI use does not have to collapse. The correction is more likely to arrive through higher prices, rationed capacity, provider consolidation, reduced quality or credit losses. There are significant sovereignty implications for Canada.
TREND: The AI Value Stack Is Being Reordered
Put the week’s developments together and a new hierarchy begins to emerge:
| Layer | Sources of value and scarcity | This week’s signal |
| Physical infrastructure | Energy, chips, data centres, networks and capital | Nvidia financing and CoreWeave spending |
| Intelligence | Models and inference | Meta’s open-weight strategy and expanding model competition |
| Control | Data, orchestration, distribution, agents, permissions and verification | MCP, the OpenClaw incident and new government requirements |
Frontier capability will continue to matter. Genuine breakthroughs will remain enormously valuable. But as capable alternatives multiply, the economic value of each incremental improvement may decline for many ordinary tasks.
When five models can perform a job sufficiently well, the strategic question is no longer which one scored two points higher on a benchmark. What matters is which one fits most effectively, safely and economically into the surrounding system, issues that remain either unaddressed or gaping holes. That is what commoditization looks like.
THE SCUTTLEBUTT: Washington’s AI Fight Gets Personal
And then there is Dean Ball. Ball helped shape the Trump administration’s AI policy before leaving government and joining OpenAI as head of Strategic Futures.
Reactions to his posts on X this week suggest senior administration figures were deeply unhappy about the appointment. Pentagon AI chief Emil Michael reportedly went considerably further, describing Ball as the AI world’s “supreme village idiot.” Further post suggested that Ball’s role in the administration’s AI policy had been exaggerated. He then produced receipts demonstrating that his role had in fact been integral and garnered accolades.
(Source: Andrew Curran on X)
Trump administration AI policy has once again shown, after the Anthropic rug pull, which did so much damage to American AI, continues, that it has entered its messy Trump vindictive interpersonal era. The drama matters for reasons beyond its limited entertainment value. OpenAI, Anthropic, Meta, Nvidia and other AI companies increasingly depend on governments that are simultaneously customers, regulators, investors and national-security partners.
A personnel move that once might have been ordinary Silicon Valley talent competition can now become a political or geopolitical signal. Washington is no longer merely regulating the AI industry. It has become part of it, with all the rancour and public nastiness that involves.
WHAT WE’RE WATCHING
The Chinese model ecosystem: DeekSeek V4 Pro is now out, and Qwen3.8-Max (2.4T params, 95B active) dropped today as well. Models are being released at an incredible pace. Further advances from Chinese developers will undoubtedly intensify model competition, push prices lower and reinforce the argument that capable intelligence is becoming less scarce. The weights are open, they’re very good, and American AI is playing catchup. The question being asked on X after some still unvalidated DeepSeek performance numbers were shared: “Are we ready for a post-Fable-open-source world?”
OpenAI’s Astra and other new models: The rumoured next OpenAI product apparently (according to leaked documents) will be charged as “intelligence by the minute.” That is a meter and that is new, and nobody seems to be talking about tokens anymore … Grok 4.6 is now out, 4.7 and Fable 5.1 are also rumoured to be coming soon.
Meta: Muse Glimmer is one release, but can Meta can turn open-weight intelligence into a durable platform and distribution strategy? Given their historical focus on the consumer market and their recent lack of success moving products to market (virtual reality and what looks so far like a stalled AI enabled glasses rollout), plus high profile departures and a lack of adoption by the corporate market, Meta is hungry for its next success.
Nvidia and CoreWeave, Debt and Demand: Demand remains strong. Financing costs, utilization, margins and returns will determine whether the infrastructure boom is sustainable, and who benefits. Nvidia’s position seems unassailable, but if we’ve learned anything in this industry it’s that innovation can change trajectories in a matter of weeks. CoreWeave’s debt situation as AI becomes increasingly commoditized is a major liability to watch.
Agents Gone Wild? As Dr Dorothea Bauer recently pointed out on X, *all* of the disclosures about “rogue” agents have come from the companies themselves, not from whistleblower. This is a reversion of normal alarm and corporate control. Why is it happening this way? What do these outbreaks really signify? why are the companies themselves so willing to broadcast these failures perhaps because they do not actually see them as failures, but as intimidating/motivating advancements of the technology, when in fact, they are just further demonstrations of their architecture
Washington and the EU: Watch federal procurement, national security and efforts to pre-empt state AI regulation, and continued usage and disclosure policy changes throughout the ecosystem due to the EU AI Act.
The (inverted) bubble: Falling demand is not happening. Wider distribution and commoditization however could have the same market effect. There are more than one possible failure modes. Meanwhile, DeepSeek and Kimi K3 are raising prices. Debt, financing costs, power commitments, utilization rates, competition, innovation and return on capital may reveal more, as will rumoured pricing model changes.
THE BOTTOM LINE
The AI industry spent years trying to manufacture intelligence. It succeeded well enough to change where scarcity sits. Capable intelligence can increasingly be bought cheaply, downloaded, routed between providers, embedded in applications and connected to tools.
The premium is moving to the compute that runs it, the data that grounds it, the protocols that connect it, the permissions that constrain it, the verification systems that establish trust and the distribution channels that determine whose intelligence gets used.
The current phase of generative AI is a race to create intelligence. The next will be a race to control what increasingly abundant intelligence can reach, change and become, and do, under what regulation.
