AI moves onto devices as markets ask how it pays off – October 9 daily briefing

Today I’m looking at two changes at once: more AI work can run on the device in your hand, while markets are asking how the industry will turn its large investments and expectations into revenue. Investment comments are personal observations, not recommendations to buy or sell.

Information cutoff: October 9, 2026, 8:30 a.m. Korea time. I am not repeating yesterday’s Windows local-AI strategy, South Korea’s K-GX plan, the September FOMC minutes, Google’s power agreement, or the FinCEN withdrawals. GitHub’s secret-detection announcement is dated October 7, so I have left it out of this issue’s “new since yesterday” selection. GitHub’s announcement. Today’s four items are a new GPU runtime, Samsung’s official preliminary results, a difference in how OpenAI revenue is counted, and fresh Fed and oil signals.

1. Google’s ML Drift: a translator between AI models and device GPUs

On October 8, Google released ML Drift, an open-source GPU runtime for on-device AI under the Apache 2.0 license. A model may be the same, but phones and computers use different GPUs, drivers, and programming interfaces. Developers often have to adapt code for each one. ML Drift aims to connect those layers through a shared execution system that works with interfaces such as OpenGL ES, OpenCL, Metal, and WebGPU. In simple terms, it can take some of the translation work between an app’s model and the device’s graphics chip off the developer’s hands. Google Developers’ ML Drift announcement

The runtime powers LiteRT’s GPU accelerator and is also available as a standalone C++ library. Google says the older TensorFlow Lite GPU delegate will no longer receive new feature updates and recommends migrating to ML Drift. The desktop paths are at different stages: Windows and Linux use WebGPU in an early preview, while macOS uses Metal. So I would not read “cross-platform” as meaning every system is already equally mature.

Google reports that ML Drift cut average frame latency for YouTube Shorts effects by up to 40%, made selected photo-editing and Snapchat features up to 30% faster, and reduced memory overhead by up to 12% in its Gemma tests. Those are Google’s own measurements, not independent guarantees across every device. Still, the basic benefit is easy to see: a quick camera effect can be processed on the phone instead of making a round trip to a server, which can reduce delay and data transfer.

The important change for me is the execution layer, not a new model. Yesterday’s Windows announcement focused on the operating system and the permissions agents receive; ML Drift works lower down, between a model and a device’s GPU. These are different layers, not duplicate features. Developers still need to test compatibility, heat, battery use, device-to-device speed differences, and unsupported operations. I expect local AI to complement cloud AI: small, immediate jobs can stay on the device, while larger reasoning tasks still go to the cloud.

2. Samsung reports KRW 107.4 trillion in preliminary Q3 operating profit—and its shares fall

On October 8, Samsung Electronics reported preliminary third-quarter figures of KRW 195 trillion in revenue and KRW 107.4 trillion in operating profit. Operating profit was up 782.5% from a year earlier and 20.01% from the previous quarter. These are official preliminary figures provided before Samsung finishes closing its accounts; they are not a final report with confirmed segment-level details. Samsung’s preliminary Q3 results

The figure was a little above the KRW 106.1 trillion estimate reported the day before. Yet Samsung shares fell 2.4%, and the KOSPI fell 2.6%. U.S. technology and chip stocks were also weak, so it would be too simple to attribute a full day’s decline to Samsung’s earnings alone. Reuters’ prior consensus estimate, AP’s October 8 market report

I do not read this as “the results were bad.” The preliminary profit was enormous, but investors may have been looking past one quarter to ask whether that pace can continue. A single day’s share price cannot prove that earnings have peaked either. Until Samsung publishes its full results and more forward-looking detail, I think it is safer to separate the record number from the market’s expectations for what comes next.

3. OpenAI’s reported $50 billion run rate: a difference in accounting, not a sudden $20 billion drop

Reports on October 8 said OpenAI had told investors that its annualized revenue was approaching $50 billion at the end of September. That looks far below the roughly $70 billion figure reported earlier, but it does not mean that OpenAI suddenly lost $20 billion in revenue or formally cut its guidance. An annualized run rate projects a recent sales pace over a year. According to the reports, the two figures treated sales through cloud partners differently. The earlier $70 billion figure was close to a comparison adjusted to include partner sales in a way more like Anthropic’s reporting. Axios on the accounting difference, Reuters’ report

What stood out was how sensitive markets were to the report. On October 8, the Nasdaq fell 1.3%; Nvidia dropped 2.9%, Broadcom 4.3%, and Micron 4.8%. Oil, interest rates, and geopolitical concerns were also moving markets that day, so the OpenAI story cannot be treated as the sole cause. AP on technology stocks and the market, U.S. market close

Token usage or a large run-rate headline is not enough to tell whether an AI business is healthy. We also need to know what portion of partner sales the company records as revenue, what inference costs, what long-term compute commitments it has, and how much cash it generates. The question I see widening is: “Will GPU orders keep growing?” to “Can the companies selling AI services earn recurring revenue and profit to support the investment?” This report does not settle that question, but the market has begun reacting to it.

4. A Fed official sees possible tightening over six to nine months; Brent reaches $104

On October 8, St. Louis Fed President Alberto Musalem said that further rate increases could be needed over the next six to nine months to bring inflation back toward target. He did not say the Fed should raise rates at its October meeting; he said he had not prejudged that decision. The St. Louis Fed’s own page confirms the event and its subject, while Reuters reports the specific six-to-nine-month comment. One regional Fed president’s remarks are not a firm policy path for the whole committee. St. Louis Fed event note, Reuters on Musalem’s remarks

Brent crude rose 4.1% to $104.28 a barrel on the same day; U.S. WTI rose 3.6% to $91.49. Middle East supply risks were joined by Hurricane Isaias approaching the U.S. Gulf Coast, where roughly 1.3 million barrels per day of offshore oil production had been temporarily shut in. Oil ended below its session high, and the 10-year Treasury yield also eased to 5.221% after an auction—another reminder that markets do not always move in one direction at once. Reuters on oil and supply disruptions, U.S. Treasury and technology-stock close

My reading is that this was a renewed warning that rates could stay high if inflation proves persistent, not a decision to raise them immediately. A few days of higher oil prices do not by themselves mean long-term inflation is entrenched; the duration of supply disruptions and any pass-through into transport and other prices matter. Still, when borrowing costs and energy prices are both unsettled, businesses that spend heavily now and expect returns much later can face a tougher test.

My take today

Today’s news comes back to two questions: “Where does AI run?” and “How does the investment get paid back?” ML Drift widens the path for AI to run on devices, but performance and compatibility still need to be tested on real hardware. Samsung delivered record preliminary earnings, yet expectations shaped the share-price response. The OpenAI revenue debate shows investors looking beyond headline usage and asking about accounting and cash generation. What I will watch next is not only model performance, but the cost of real use, recurring revenue, and whether the business can hold up as power and financing costs change.

This is a personal record for understanding markets and policy, not a recommendation to buy or sell any asset. Each reader is responsible for their own decisions.