Nvidia used to sell chips. Increasingly, it finances the companies that buy them, guarantees the infrastructure that houses them, invests in the power and land required to run them, funds the models that consume them and helps create financial instruments through which everybody else can invest in the resulting infrastructure.

At some point, the world’s most valuable semiconductor company started looking surprisingly like a bank. This week made that transformation difficult to ignore.

Nvidia is working with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on financing platforms intended to mobilize more than $500 billion in third-party capital for AI infrastructure. Nvidia itself describes the goal as turning compute and AI infrastructure into an “investable asset class.”

Then came Ohio. Nvidia agreed to provide financial support potentially reaching $105 billion for SB Energy’s enormous PORTS-Pike data-centre development, where OpenAI has signed a 20-year lease. Nvidia also invested $1.5 billion directly into SB Energy. The first phase alone covers 4.25 gigawatts of IT load, with the potential campus reaching 8 GW.

Nvidia has also moved further upstream, investing in companies securing power and data-centre sites. And now it is moving downstream too.

MACRO: The Bank of Nvidia

The most interesting thing about Nvidia today may no longer be its chips. It is the ecosystem Jensen Huang is constructing around them. The basic problem Nvidia faces is almost comically fortunate: demand for AI compute is enormous, but customers need staggering amounts of money, electricity, land and infrastructure to buy as many Nvidia systems as they want.

Nvidia’s solution increasingly appears to be: help them get all of it. The company isn’t lending $500 billion itself. The new financing platforms are designed to bring enormous pools of institutional capital into AI infrastructure. But Nvidia increasingly occupies the centre of the transaction.

It supplies the chips and the software. It invests in customers and adjacent companies, helps secure land and power, provides residual-value support for infrastructure containing its equipment and benefits when all of those activities produce additional demand for Nvidia products.

S&P Global has already started accounting for Nvidia’s residual-value guarantees as a debt adjustment. It kept Nvidia’s AA rating stable, but specifically warned that increasing leverage throughout the AI ecosystem could create greater volatility for Nvidia over time.

This is why the circular-financing criticism can’t simply be dismissed. It doesn’t mean demand is fake. Quite the opposite: demand for compute remains extraordinary.

But Nvidia is increasingly using the wealth created by that demand to finance the infrastructure necessary to generate more demand. That is an extraordinarily powerful flywheel while it works. It also means Nvidia is acquiring exposure to considerably more than semiconductor sales.

MACRO: Nvidia Starts Financing the Model Layer

Then came Poolside.

Nvidia is reportedly paying approximately $6 billion for a non-exclusive licence to Poolside’s model-development technology, investing another $1 billion in the company and bringing more than 100 Poolside engineers into Nvidia’s Nemotron effort. The objective is explicitly to strengthen an American open-weight alternative to increasingly capable Chinese models.

That is a remarkable change in Nvidia’s role. The company that supplied the hardware for the model race is increasingly participating in the model race itself. But there is an irony here. The open-weight movement began partly as an ideological proposition: intelligence should not be controlled by a tiny collection of proprietary labs. What Nvidia is constructing looks much more commercial.

Open weights are valuable because they increase model deployment. More deployment increases inference. More inference consumes compute. And Nvidia sells compute.

There is nothing inherently wrong with that. In fact, it may be the strongest economic argument for open weights in the American market. But it is distinctly different from the original idealism surrounding open AI (no pun intended). Open has become a business strategy. Nvidia does not need Nemotron to become ChatGPT. It needs thousands of companies to deploy models.

Its own technology is also accelerating. On August 24, Nvidia said its Groq 3 LPX inference accelerator is now in full production as part of the Vera Rubin platform. Nvidia is positioning it specifically around the inference demands of agentic systems, where long contexts and repeated token generation make latency and throughput increasingly important. In an Artificial Analysis benchmark using Gemma 4 31B at 100,000-token context, Nvidia says LPX delivered 3,400 output tokens per second, roughly 4× the nearest alternative platform.

MACRO: And Now Perplexity

Nvidia is also reportedly discussing an investment in Perplexity at a valuation above $30 billion, up more than 50% from its previous financing valuation.

The interesting number isn’t actually the valuation. Perplexity’s annualized revenue has reportedly grown from below $250 million at the beginning of 2026 to more than $750 million, driven substantially by Perplexity Computer, its agentic professional-work product.

This is where Nvidia’s strategy becomes clearer. It doesn’t have to know which AI company wins. It can own pieces of many of them.

OpenAI needs infrastructure? Nvidia helps finance it. Open models threaten proprietary labs? Nvidia funds open models. Perplexity finds an enterprise niche? Nvidia can invest there too. Data centres can’t find power? Nvidia can invest in the infrastructure layer.

The strategy increasingly resembles a bank crossed with an industrial conglomerate: finance the ecosystem, supply the productive asset and participate in the upside. The Bank of Nvidia doesn’t need to pick one winner. It needs AI usage to keep growing.

MACRO: China Is Selling a Better Life. America Is Selling the Apocalypse.

There is another increasingly striking divergence in AI that has little to do with benchmarks. It is how the technology is being communicated.

In much of the Chinese presentation of AI, the message is remarkably straightforward: AI will make life better. It will improve factories, help doctors and make cities more efficient. Robots will perform dangerous work. AI will compensate for an aging population. Productivity will rise and new industries will emerge.

The government’s “AI Plus” strategy explicitly encourages the diffusion of AI throughout the economy, and even reporting documenting genuine anxiety among Chinese workers about displacement finds considerable curiosity and optimism alongside it.

Compare that with the dominant American conversation: AI may eliminate your job, destroy creative industries, manipulate elections, escape, become uncontrollable, cause catastrophic harm or eventually kill everybody. Some of those concerns are legitimate. We have covered many of them here. But communications matter.

China is largely selling AI as industrial modernization. America frequently communicates AI as an extraordinary technology it is simultaneously racing to build and terrified of building. Europe often adds another layer: something potentially useful that must first be made safe enough to permit.

These aren’t merely differences in propaganda. They can influence adoption. A population taught to associate AI with economic improvement may react differently to deployment than one repeatedly told the technology threatens its employment, culture, privacy and perhaps existence.

There is an obvious risk in China’s framing too. Optimistic state narratives can obscure displacement, surveillance, labour disruption and failures. But the communications asymmetry is becoming difficult to ignore. The countries competing most aggressively to build the AI future are telling their populations radically different stories about why they should want it.

MICRO: Who Is the Mystery Model?

Something strange is happening on the model leaderboards again. Anonymous models have appeared in public testing, generating the now-familiar ritual in which researchers and AI obsessives attempt to reverse-engineer their identities from behaviour, style, capabilities and benchmark performance.

This has become one of the peculiar institutions of frontier AI. Arena testing deliberately hides model identities during comparisons, allowing users to evaluate outputs without knowing which company produced them.

Because major labs increasingly test unreleased systems this way, an unidentified strong performer can trigger an industry-wide guessing game. Is it OpenAI? Google? Anthropic? Z.ai? Alibaba? Something nobody is expecting?

There may now be more than one interesting mystery model circulating. The phenomenon, while largely a marketing tactic, itself says something about how compressed the frontier has become. A year or two ago, model families often had recognizable capability gaps. Increasingly, a genuinely strong anonymous model can plausibly belong to any of several American or Chinese labs. That uncertainty is itself evidence of competition.

MICRO: The Model Conveyor Belt Speeds Up

The mystery models are arriving into an already crowded market. DeepSeek is iterating V4. Qwen continues expanding. Kimi remains competitive. Z.ai has new GLM systems coming through testing. MiniMax is iterating. Mistral is repositioning. Nvidia is pouring billions into Nemotron and Poolside.

The important development isn’t which benchmark moved three points this week. It is that major model releases are becoming continuous rather than episodic.

The model market increasingly resembles software development. There is always another release, another preview, another open-weight alternative and another cheaper inference option.

That has profound implications for model economics. The useful life of a capability advantage is shrinking. Which brings us directly back to Nvidia: if models become increasingly interchangeable, the company selling the infrastructure underneath all of them may occupy the safest position in the market.

POLICY: Canada-US Trade Talks Collapse, and AI Sovereignty Gets More Real

Canada and the United States spent much of the week attempting to salvage their trade negotiations.

As the world is now well aware, Canada walked away after what Prime Minister Mark Carney described as unacceptable American demands affecting Canada’s economic independence and sovereignty. The United States subsequently imposed 50% tariffs on roughly $20 billion of Canadian goods, with Canada announcing retaliation beginning September 8. This has implications for Canadian AI policy.

Canada’s AI infrastructure is still deeply integrated with the United States. Cloud providers are American. The dominant frontier models are overwhelmingly American. Much of the semiconductor stack is American-controlled. Major enterprise AI platforms are American. Canadian defence and public-sector systems increasingly depend on American technology ecosystems.

For decades, that dependence was easier to justify because economic integration with the United States was treated as essentially permanent. That assumption is fractured.

Carney’s own statement following the breakdown emphasized that Canada “will not return to our old relationship” and must build greater domestic strength while diversifying partnerships abroad.

For AI sovereignty, that shifts the policy question. Sovereignty increasingly means understanding which dependencies Canada can tolerate when the political relationship underlying those dependencies less and less stable, and how much those dependencies are tolerable, what workarounds are available, and what t investments are necessary.

That doesn’t mean technological autarky. Canada cannot and should not recreate the entire AI stack domestically. But the trade dispute makes redundancy, domestic capacity, vendor diversity, data jurisdiction and control over critical public-sector systems look considerably less theoretical.

GOSSIP: China Held the Robot Olympics and Everyone Else Had Feelings

And then China held the World Humanoid Robot Games. More than 2,000 robots from 16 countries competed in 51 events in Beijing, including running, jumping, football, martial arts and an increasingly important set of practical tasks involving factories, restaurants and other real-world environments.

Some of it was objectively hilarious. Robots ran spectacularly fast and then discovered that stopping was a separate engineering problem. There were crashes, padded walls and robot sporting events that looked considerably less threatening than their promotional videos suggested. But laughing at the robots misses the point. The Games are an enormous iterative testing environment.

More than 40% of the practical tasks reportedly require full autonomy. Robots are being tested on plugging in cables, charging vehicles, manipulating objects, recognizing environments and recovering from mistakes. And China is doing this at scale. That is the part producing visible anxiety in American and European technology circles.

The United States pioneered enormous amounts of modern robotics research. Reuters documented just this week how U.S. military-funded research helped produce techniques subsequently commercialized extremely effectively by Chinese robotics companies such as Unitree.

China’s advantage increasingly appears less mysterious. It is very good at turning research into manufacturing: build robots, build lots of them, make them cheaper, put them into competitions, crash them, fix them, generate data and repeat.

There are legitimate questions about how much genuine commercial demand exists. Government-backed training centres themselves buy significant numbers of Chinese humanoids, creating a degree of circularity in the market. But industrial policy is doing what industrial policy is designed to do: accelerate learning curves before the market necessarily exists.

The robot falling into the wall is funny. The factory that can manufacture another 10,000 improved versions of it is considerably less funny if you’re competing against it.

WHAT’S COMING: AUGUST 25–31

GLM-5.3 open weights: This is the only major model release with anything approaching a firm timetable. Z.ai launched restricted access to GLM-5.3 on August 14 and said the weights would follow approximately two weeks later, after additional security testing. That puts the likely release around August 28. The important question is whether Z.ai releases the complete cyber-capable model or restricts some functions through its proposed trusted-access system. Z.ai

Nvidia earnings — August 26: Watch less for another enormous revenue number than for what Nvidia says about the financial architecture surrounding that revenue: residual-value guarantees, infrastructure commitments, customer investments, China sales and the amount of risk accumulating beyond ordinary semiconductor sales. Nvidia Investor Relations

Marvell earnings — August 27: Marvell sits in the less visible but increasingly important layer of the AI infrastructure market: networking, custom silicon, interconnects and memory infrastructure. Its results should provide another signal about whether AI spending is broadening beyond GPUs into the systems required to connect and operate them. Marvell Investor Relations

Ray Summit — August 24–26: Anyscale’s developer conference is focused this year on foundation-model training, multimodal data, reinforcement learning and physical AI. Nvidia is scheduled to discuss how it is building and post-training Nemotron with Ray, while the co-located vLLM conference will focus on serving increasingly enormous open models efficiently. Watch for technical disclosures about Nemotron, inference economics and the infrastructure required to train agentic systems. Ray Summit

NOAA AI Workshop — August 25–27: NOAA’s three-day workshop in Boulder will concentrate on AI weather and climate emulation, machine-learning forecasting and environmental science. It is unlikely to produce a consumer product announcement, but it may offer a useful view of where scientifically grounded public-sector AI is actually moving into deployment. NOAA

AI in Healthcare — August 26–28: Imperial College London hosts the third International Conference on Artificial Intelligence in Healthcare, with research spanning medical imaging, clinical decision support, diagnostics and responsible deployment. Watch for evidence that medical AI is progressing beyond retrospective benchmark performance toward validated clinical use. AIiH 2026

Shenzhen’s AI and robotics exhibitions — August 26–28: AGIC and the enormous IOTE technology exhibition will bring together foundation models, digital humans, embodied intelligence, industrial robots, drones and factory applications. After the Humanoid Robot Games, this may provide a better indication of which Chinese robotics systems are becoming commercial products rather than demonstrations. IOTE Shenzhen

California’s AI legislative deadline — August 31: California lawmakers have until the end of the month to pass roughly two dozen remaining AI-related bills. Measures awaiting final action include restrictions on AI “robobosses,” rules governing lawyers’ and arbitrators’ use of generative AI, expanded provenance requirements, healthcare AI restrictions and protections for children interacting with chatbots. What survives this final week could become the next de facto national technology standard. California Senate legislative calendar

LEAP and DeepFest — beginning August 31: One of the world’s largest technology events opens in Riyadh at the very end of the week. Given the newly announced Mistral-HUMAIN partnership, watch for additional sovereign-AI agreements, infrastructure commitments, Arabic-language model development and Saudi investments across the model and compute stack. DeepFest

No major American lab has formally scheduled a frontier-model release for the week. Nor has DeepSeek, Qwen, Kimi or Mistral provided a confirmed date for another new model. GLM-5.3’s weights are the one reasonably well-defined release to watch. Everything else, including the identities of the mystery Arena models, remains possible, but unscheduled.

THE BOTTOM LINE

The strangest (and strongest) AI company in the world right now is Nvidia.

It makes the chips everybody wants, then invests in companies that use them. It helps finance the data centres where they run, supports the residual value of the infrastructure containing them and invests in the companies securing the power and land. It funds open models that create additional inference demand. And now it may invest in the applications sitting at the very top of the stack.

The Bank of Nvidia is financing an economy in which Nvidia sells the principal productive asset. That may be brilliant. It may create uncomfortable circularities. It is probably both.

And while America builds increasingly sophisticated financial structures around AI, China is doing something almost comically tangible: building things. Models, robots, factories and infrastructure, then building more of them.

Perhaps the most important contrast in AI this week isn’t between American and Chinese benchmarks. It is between two different theories of how technological leadership is created.

One is increasingly financial. The other is increasingly industrial. Both are now moving extraordinarily fast.