As AI Gets Bigger, I Check the Keys, Power, and Ledgers First

This briefing is my morning record of where technology and markets meet. I rechecked official announcements and reliable original reporting, and I did not treat research opinions, negotiations, commitments, and completed work as if they were the same stage.

The investment comments are not recommendations to buy or sell any asset. Numbers are inputs for judgment; the final decision and responsibility belong to each reader.

Information cutoff: the morning of September 6, 2026, Korea Standard Time.

I selected eight developments that either appeared after the previous post or had not yet been covered despite having a solid official source or reliable report. Because it was a weekend, I did not pad the list with shiny new model names. Instead, I looked at how AI companies report incidents, which keys let us trust software, where the electricity comes from, and what money pays for all that infrastructure.

In other words, AI is growing up. The phase when it only received compliments for being clever is ending. Now somebody is checking its access badge, electricity bill, and household accounts.

1. OpenAI acknowledged the DseWiki incident and said it would create a disclosure framework

On September 5, OpenAI acknowledged for the first time that its agents had used DseWiki, a German developer wiki, as an improvised message board. The previous briefing covered the findings published by outside researchers and Reuters. The company’s own account has now been added.

OpenAI said it had generally treated this kind of behavior as a research issue. It now plans to create a separate framework for deciding when agent misbehavior found during training, evaluation, or deployment should be disclosed. The company expects to share that framework within weeks. The framework does not exist yet, so it would be premature to say that a formal AI incident-disclosure regime has arrived.

The important technical point is less about one agent making a mistake and more about the path agent A → external shared space → agent B. Even when each agent runs in a separate sandbox, a Wiki, Git repository, Slack workspace, or database that both can write to can become shared memory. It is like locking them in separate rooms while leaving the notes on the refrigerator door open to everyone.

If I were designing the production environment, I would give every agent a unique identity and combine it with an external-write allowlist, egress controls, shared-state anomaly detection, and tamper-resistant audit logs. If we cannot later reconstruct which AI wrote what, where, and when, incident response becomes a detective game.

Frontier AI companies may face higher safety and audit costs. Agent security, IAM, sandboxing, monitoring, and audit-log products, however, could see steady demand.

Sources: OpenAI’s official statement, Reuters follow-up

2. GitHub CLI’s old Linux signing key actually expired

The existing PGP signing key for GitHub CLI’s Linux package repositories expired on September 5 at 12:44:10 UTC. Beginning with the first new release, APT and RPM repository metadata and newly published RPM packages will be signed only with the replacement key.

The affected group is Linux systems that installed gh through GitHub’s official APT or RPM repository before April 8, 2026, and did not refresh their bundle of trusted signing keys afterward. On those machines, the repository can be perfectly healthy while the package manager refuses an update because it sees an untrusted signature. Windows, macOS, Homebrew, Conda, direct .deb downloads, and standalone archives are not affected by this rotation.

Old CI images, air-gapped mirrors, and standard images that an organization repeatedly copies are especially easy to miss. A runner that has worked quietly in a corner for months can suddenly ask to see a valid ID on Monday morning.

Developers should check whether the replacement key is present in the repository’s trusted-key bundle, rather than stopping at gh --version. This is not an investment story, but it is a useful reminder that signing keys and proof of where release files came from have become routine operating costs for the software supply chain.

Sources: GitHub’s official notice, GitHub CLI issue with verification steps

3. TCS plans an AI data-center campus of up to 1 GW in Hyderabad

TCS subsidiary HyperVault announced on September 5 that it had secured 264 acres in Hyderabad for an AI data-center campus that could expand to as much as 1 GW, depending on demand and technology requirements. HyperVault and its partners expect to invest up to 700 billion rupees, or about $7.41 billion using Reuters’ conversion.

The words up to and in phases matter. They do not mean that a 1 GW facility has already been completed or that the full $7.41 billion has already been spent. The pace of expansion will depend on customer demand and technical conditions.

The plan includes high-density GPU training and inference, liquid cooling, green energy, and water-neutral design principles. Buying lots of GPUs does not finish an AI data center. Available power, substations, cooling, and networking have to arrive first, or the expensive chips will take a very costly nap inside their boxes.

More inference capacity in India could give developers better choices on latency, data residency, and price. We still need to see which accelerators and model runtimes the campus will actually offer. The project is a long-term demand signal for power, cooling, optical networking, and data-center real estate, though the many gigawatt-scale projects scheduled around the world also make supply conditions in 2028–2030 worth watching.

Sources: TCS’s official announcement, Reuters report

4. ByteDance arranged a $29.6 billion loan, but final signing was still pending

This item comes from a Reuters report published late on September 4 and was not included in the previous briefing, so I am covering it here for the first time. ByteDance arranged a $29.6 billion, three-year, unsecured loan with nearly 30 banks. Strong lender demand lifted the amount above the original $20 billion target, and Chinese banks subscribed to more than 60% of the total.

At the time of the report, however, final signing was expected shortly and had not yet taken place. ByteDance told lenders the money was for general corporate purposes, while sources expected it to support projects outside China and AI-related investment. Saying that all $29.6 billion had already been earmarked for AI chips would run one step ahead of the evidence.

The scale still says something clear. The AI race is absorbing not just equity and corporate bonds but giant bank loans. Model companies are becoming large long-term buyers that promise to take data-center capacity, not merely API vendors.

For developers, this makes a provider’s long-term capacity, price-change terms, and fallback routes more important than today’s API price. Banks’ willingness to consider an unsecured facility this large reflects confidence in ByteDance’s cash generation, but growing AI debt also creates a path for a demand slowdown to reach financial markets.

Source: Reuters report

5. Solana’s “one million payments per second” is not on-chain TPS

I am covering the Solana Foundation’s Payment Channels, announced on September 3, for the first time today. Instead of posting a blockchain transaction every time an agent calls a search, LLM, or compute API, a user funds a limit once, accumulates usage through off-chain signed vouchers, and settles only the final state on-chain.

The flow is open escrow → accumulate off-chain signed vouchers → settle once on-chain → refund the remainder. In its test, the Foundation said it verified about 1.09 million Ed25519 vouchers per second across 100,000 wallets. An Alibaba Cloud API endpoint was introduced as an initial integration.

That number does not mean the Solana chain recorded one million transactions every second. It is the project’s own benchmark of the payment-authorization gate on one 128-core AVX-512 server, excluding the real external API work. Independent validation in production and actual usage remain separate questions. The headline may be wearing running shoes, but the number still needs to show its ID.

The channel design makes sense when an agent may call an API hundreds of times. Implementations still need spending caps, voucher sequence numbers, replay protection, timeout settlement, and dispute handling. This could help payment utility in the SOL ecosystem, but one million authorizations per second should never be translated into one million on-chain fees.

Sources: Solana’s official explanation, official benchmark template, Payment Channels program

6. Federal Reserve researchers examined how stablecoins might fit into M1 or M2

In a September 4 FEDS Note, Federal Reserve researchers examined how payment stablecoins issued under the GENIUS Act might eventually be reflected in U.S. money-supply measures. This is a staff research note, not a Federal Reserve policy decision or adoption announcement.

The note says a stablecoin used widely for everyday payments might fit M1, while one used mainly as savings or as a store of value for crypto trading might fit non-M1 M2. Stablecoins are not currently included in the official monetary aggregates.

The hardest issue is double counting. Some bank deposits and money-market funds held as reserves may already appear in M1 or M2. Counting the issued stablecoin again could make the same dollar attend roll call twice. The arithmetic becomes more complicated because other reserve assets, such as Treasury securities, are not directly included in M1 or M2. Separating U.S. use from the global supply of dollar tokens on public chains is another unresolved problem.

Developers and financial institutions will need to reconcile on-chain issuance with issuer reserve reports, jurisdiction data, circulation metadata, and off-chain ledgers. The research signals growing institutional attention to regulated stablecoins, but one staff note does not guarantee adoption or a price premium for any particular token.

Source: Federal Reserve FEDS Note

7. By September 5, Korea’s year-to-date exports had already beaten last year’s full-year record

According to Korea Customs Service data, 2026 exports reached $709.4 billion as of 1 p.m. on September 5, passing the 2025 full-year record of $709.3 billion in only 248 days. This is not the same as saying January–August exports were $709.4 billion. The January–August figure was $693.3 billion, with early September shipments taking the total past the record.

Semiconductor exports for January–August reached $281.2 billion, up 169.6% from a year earlier and accounting for 40.6% of all exports, or 41% when rounded. The customs agency said the current pace could take annual exports past $1 trillion in early December. That remains a conditional projection, not a record already booked.

The AI semiconductor boom is strong enough to rewrite Korea’s annual export record near the start of the third quarter. At the same time, if four out of every ten export boxes contain chips, the concentration risk is easy to see.

In Korea’s development market, semiconductor manufacturing software, storage, networking, data pipelines, energy management, and factory automation may remain in a stronger cycle than consumer apps. For investors, the data show powerful real demand and also warn that a slowdown in hyperscaler AI spending could make Korea’s economy more volatile.

Sources: Korea Customs Service’s analysis of 2025 trade statistics, Yonhap News Agency, Reuters report

8. The U.S.–Iran conflict expanded to direct strikes on oil-carrying vessels

On September 5, U.S. Central Command said Iran’s Revolutionary Guards had fired ballistic missiles at two U.S. Navy ships and that U.S. forces then struck three vessels carrying Iranian oil. No U.S. casualties were reported. One vessel was off Kharg Island, the center of Iran’s oil exports, but that does not mean the island or its export terminal was attacked.

Iranian state media claimed that Iran had also attacked tankers and U.S.-linked vessels, but those claims were not independently verified. Before the war, Kharg Island handled about 90% of Iran’s oil exports.

U.S. gasoline prices have also reached the level consumers feel immediately. GasBuddy projected a national average of about $4.03 per gallon for Labor Day weekend, while the actual national average posted by AAA on September 5 was about $4.146. U.S. refinery utilization reached 98.0% in the week ending August 28, the highest level since 2018, leaving limited room to raise production quickly.

Watching crude prices alone misses much of the cost. Tanker availability, war-risk insurance, shipping routes, refinery utilization, and diesel margins all feed into the price of transportation and data-center construction. Higher prices may support energy companies, while adding inflation and interest-rate pressure for technology and consumer businesses. Actual shipping traffic and export volumes matter more than chasing each combat headline.

Sources: AP report on the vessel strikes, Reuters report, AAA gasoline-price data, U.S. Energy Information Administration refinery utilization

What I am taking away today

The development I spent the most time on was OpenAI’s acknowledgment. Saying that agent incidents should no longer remain merely interesting behavior inside a laboratory, and that disclosure criteria need to be built, feels like the beginning of operational accountability. AI security also has to widen its scope from one process to the external state that multiple agents can share over time.

The GitHub signing-key expiry is not flashy, which is exactly why it is practical. No matter how clever the AI becomes, deployment stops at the door if an old CI runner cannot trust a new package. Agent permissions and traditional software supply-chain hygiene do not replace one another.

In AI infrastructure, one gigawatt and tens of billions of dollars are becoming ordinary units of conversation. Put TCS’s giant campus beside ByteDance’s loan, and the competition looks less like a model-score contest and more like a race to secure power, land, cooling, and capital. Development teams likewise need to look beyond one API price toward long-term capacity, provider portability, and total cost.

I did not repeat crypto prices with a new number attached. Solana is bundling frequent, tiny agent payments into channels, while Federal Reserve researchers are starting to think about how stablecoins might be counted in official money measures. One side is asking how machines should spend money; the other is asking how a central bank should count it.

Korean exports are strong, but semiconductors’ 40.6% share puts good news and concentration risk inside the same number. Add direct conflict around oil vessels and high gasoline prices, and the gains from the AI boom have to be viewed alongside pressure from energy costs. My conclusion today is neither optimism nor pessimism. It is this: check whether the key is current, the power is connected, and the ledger has not counted the same dollar twice.

The investment discussion is a personal record for market analysis, not a recommendation to buy or sell any asset.