Models Changed, Capital Connected

The supplied briefing is part of the material I use to keep up with changes in the fields I follow. For this public post, I retained only facts I could verify again through official announcements and reliable original reporting.

Investment comments are my market notes, not advice to buy or sell any asset. For numbers that move with time, such as annualized revenue and rate probabilities, I state both the measurement and its timing.

Information cutoff: 8:53 a.m. KST on September 1, 2026, including the August 31 US close.

The AI industry looked like a giant transfer station today. Six models got off the GitHub platform, Nvidia built a $3.5 billion passage to MediaTek, and OpenAI put an advertising checkout beside the conversation window and printed a $1 billion annualized number. The passengers change often; the companies that own the tracks and turnstiles tend to stay.

My biggest takeaway is that connecting interchangeable models, chips, funding and distribution now matters more than choosing one excellent model. Models change like seasonal menus, while data centers and bonds are long-term commitments. The AI refrigerator changes every week; the financing plan can last decades.

1. Nvidia put a $3.5 billion link around MediaTek

Nvidia invested $3.5 billion in overseas convertible bonds issued by MediaTek. Nvidia took most of the $3.9 billion offering, and Alphabet also participated. This looks less like a passive bet on the share price and more like a strategic investment: MediaTek will adopt NVLink Fusion, helping customers connect custom XPUs to Nvidia rack-scale systems.

In plain language, customers may design their own AI chips while still driving on Nvidia’s data-center highway. Custom ASICs and Nvidia GPUs do not necessarily form a winner-takes-all choice. Heterogeneous systems that mix several accelerators on one fabric are likely to become more common.

For developers, the question shifts from CUDA or ASIC? to how naturally the serving layer can operate across accelerators. For investors, Nvidia is extending its moat from GPUs into interconnects and system architecture. The caution is circular financing: as a supplier keeps funding ecosystem partners, it becomes harder to tell whether investment follows demand or helps manufacture it.

Sources: Nvidia and MediaTek’s joint announcement, Reuters analysis of the convertible-bond investment

2. ChatGPT Ads reached a $1 billion annualized pace in under 200 days

OpenAI said ChatGPT Ads reached a $1 billion annualized revenue run rate. That does not mean $1 billion has already landed in the bank. It means the latest revenue pace would equal that amount if sustained for a year. Still, reaching it in under 200 days shows that advertising has moved beyond a side experiment.

OpenAI also opened its self-service Ads Manager to advertisers in India, Europe, the Middle East and North Africa. The company repeated that ads are labeled and kept separate from answers, that advertisers do not receive private conversations, and that advertising does not influence responses.

The sound product architecture is LLM response generation → intent signal → separate ad auction and rendering. Mixing sponsored product data directly into the prompt blurs the line between a recommendation and an ad. A conversation window may have a checkout, but the cashier should not secretly rewrite the advice.

OpenAI gains another cash-flow source to offset enormous inference costs. Google and Meta gain a new competitor targeting purchase intent expressed in conversation rather than in a search box or feed.

Source: OpenAI’s ChatGPT Ads milestone announcement

3. The EU ordered LUMI-AI because it is already turning applicants away

EuroHPC signed a €387.8 million contract with French state-owned Bull to build LUMI-AI. The system will sit beside the current LUMI facility in Kajaani, Finland, and is expected to enter service in the second half of 2027. It combines AMD accelerators and CPUs with IBM storage and Nokia networking.

The most revealing line was an EuroHPC official’s admission that some applications are currently rejected because demand exceeds capacity. Sovereign AI is no longer only a policy slogan; public compute now has a waiting list. Europe is building 19 AI Factories and 13 associated antennas.

Developers using national or public compute cannot assume an Nvidia-only CUDA environment. ROCm, HPC schedulers, distributed storage and data-export rules all matter. For investors, public supercomputers are broadening demand for AMD, storage and networking beyond American hyperscalers. Governments now take a number in the GPU waiting room too.

Sources: EuroHPC’s LUMI-AI contract announcement, Reuters on capacity shortages and the system design

4. Six GitHub Copilot models clocked out today

The date in GitHub’s advance notice has arrived. From September 1, Gemini 3.1 Pro, Claude Opus 4.5 and 4.6, Claude Sonnet 4.5 and 4.6, and Raptor Mini are deprecated across Copilot. There is a limited exception for Claude Sonnet 4.6 on annual individual subscriptions, and enterprise administrators may need to enable replacement models through organization policies.

This is more operationally important than another “new model” headline. If an agent prompt and tool loop are tuned to a particular model’s habits, changing one model ID may change behavior. It might edit more files, run fewer tests or repeat the same tool call.

Production systems should therefore manage model IDs in configuration or a gateway rather than hard-code them into business logic. A regression suite should compare task success, files touched, test pass rate, retries and cost before and after the switch. Models are becoming short-term tenants. If the furniture is welded to the wall, the engineering team cries on moving day.

Source: GitHub’s official Copilot model deprecation notice

5. The FSB called AI cyber risk finance’s most immediate AI concern

Financial Stability Board chair Andrew Bailey warned G20 finance ministers and central-bank governors that frontier AI could materially change the speed, scale and economics of cyber risk. The wording was unusually direct: for the financial system, the most immediate concern is frontier AI’s potential impact on cyber risk.

This is no longer just one bank’s IT problem. If many financial institutions depend on the same model, cloud or security provider, one outage or compromise can spread at once. Defense has to move beyond quarterly checklists toward isolation → temporary credentials → egress controls → immutable audit logs → an immediate kill switch.

Giving an agent a permanent production credential is especially risky. “The AI will probably behave” is a wish, not a security policy. Regulation may slow deployment and increase operating costs, but it creates durable demand for sandboxes, IAM, SIEM and runtime isolation.

Source: The FSB chair’s August 2026 letter to the G20

6. Russia admitted crypto as a regulated investment, not as everyday money

Key provisions of Russia’s new crypto framework took effect today. A non-qualified investor may buy approved, highly liquid cryptocurrencies after passing a suitability test, with an annual limit of ₽300,000 per intermediary. Qualified investors also take a test but can access a broader set of assets without an amount limit.

Crypto payments for domestic goods and services remain prohibited. Exporters and importers, however, can use crypto for cross-border settlement without those limits. The regulated infrastructure includes exchanges, brokers and digital repositories that record ownership rights; draft capital requirements for repositories range from ₽50 million to ₽250 million depending on their activities.

Products serving Russia need a policy engine that separates jurisdiction, investor class, permitted asset, annual limit, and domestic versus trade settlement. The framework draws a bright line between crypto as an investment asset and crypto as money. Crypto entered the building, but it still cannot sit at the domestic checkout.

Sources: The Bank of Russia’s implementation summary, draft capital rules for digital repositories

7. Chinese manufacturing reached the doorstep; services stayed outside

China’s official manufacturing PMI rose from 49.2 to 49.8 in August, beating expectations. Production at 50.4 and new orders at 50.6 returned to expansion. Equipment manufacturing reached 51.4 and high-tech manufacturing 52.9.

But the non-manufacturing activity index remained at 49.0, and headline manufacturing still sat below the 50 threshold. That is why “recovery” is too simple a label. AI, equipment and export factories are warming up while the domestic service audience remains quiet.

Developers and investors looking at China should separate export-oriented B2B demand from consumer platforms. More orders for servers and EV components do not guarantee that advertising, commerce and service spending will recover at the same speed.

Sources: China’s National Bureau of Statistics August PMI release, the official sector breakdown

8. German inflation was 2.9%; energy at 10.5% was the real story

Germany’s provisional August CPI and harmonized CPI both rose 2.9% year over year. Core inflation held at 2.4%, and service inflation eased from 2.9% to 2.8%. Energy, by contrast, accelerated from 8.3% to 10.5%.

Headline inflation rose, but services and core prices did not accelerate together. It is therefore more accurate to say that the Middle East energy shock pushed the first number higher than to declare a broad inflation restart. Final August figures are due on September 10.

For European AI infrastructure, a long-term power contract can matter more than the sticker price of a GPU. Even if chips become more efficient, higher electricity prices can prevent total inference cost from falling as expected. Data centers consume tokens, but only after consuming a great deal of power.

Source: Germany’s Federal Statistical Office provisional August inflation release

9. A 4.75% US 10-year yield puts AI’s cost of capital back on the table

The US 10-year Treasury yield reached 4.75% on August 31, its highest zone since January 2025. The Dow fell 0.7%, the S&P 500 0.3% and the Nasdaq 0.1%, while energy stocks outperformed. Oil rose after the US struck Iranian facilities, reviving inflation and tightening concerns.

A CME FedWatch-based reading before the close put the chance of a 25-basis-point September hike at roughly 66%. That probability moves in real time and is nowhere near a policy decision. The more structural point came from Fed chair Kevin Warsh, who described a shift from a “global savings glut” to a “global investment surge.” Large bond issues for AI data centers are among the investments competing with Treasuries for the same pool of capital, helping keep long-term yields elevated.

When an engineering team compares a private GPU center with cloud capacity, server prices are no longer enough. The NPV needs WACC + bond spread + power + depreciation + actual utilization. Model APIs may get cheaper, but if money itself gets more expensive, the final bill has an odd habit of standing still.

Sources: AP’s August 31 US market close, Reuters on Warsh’s remarks and AI bond demand, CME FedWatch

My bottom line today

The nine stories reduce to one sentence: models are replaced quickly, while capital and infrastructure become more deeply connected. Nvidia is pulling custom chips into NVLink, OpenAI is turning user distribution into advertising revenue, and Europe is buying public compute. At the same time, six models disappeared from GitHub on one date.

If I were building a product, I would ask three questions before asking which model ranks first today:

• Can I switch providers without rebuilding the product?
• Are agent permissions and data boundaries separated from the model?
• Does the cost of one successful task include power and capital, not just tokens?

Macroscopically, Germany’s energy inflation and the 4.75% US 10-year yield are more immediate cost variables than China’s partial manufacturing recovery. AI may raise productivity, but the factories, grids and bonds required to produce that productivity consume cash first. The race between a smarter future and compounding financing costs is unusually close right now.

Korea’s finalized August trade figures were not officially available by this article’s 8:53 a.m. cutoff, so I did not fill the space with a forecast. The next briefing will compare the confirmed result with prior expectations.

Investment commentary is a personal market record, not a recommendation to buy or sell any asset. Each reader remains responsible for their own decisions.