AI Chips Need Power, Agents Need Memory, and Dollars Cost More – October 13 Daily Briefing

The shift I see today is from building larger AI models to figuring out which chips can run them continuously, and at what cost. Monetizing agents that remember, alongside a stronger dollar, makes model performance alone a poor measure of whether the business works.

Information cutoff: October 13, 2026, morning Korea time. I am not repeating yesterday’s coverage of the reported Amazon–Decart talks, Yandex data-center attacks, World Bank crisis-aid discussions, or the Riyadh airport attack. Today I’m looking at Microsoft’s new Maia 300 plan, Meta’s latest models, the latest oil-and-dollar market reaction, and the first trading-day effects of India’s FX measures.

1. Microsoft Maia 300 aims to become a primary inference system from 2027

On October 12, Microsoft unveiled Maia 300 and said it plans to scale Azure Maia to multi-gigawatt deployments from 2027, making it a primary inference system for Microsoft products and services. The company says a Maia 300 rack combines 96 accelerators, about 167 kW of power, 27.6 TB of HBM, and 1.7 PB/s of aggregate memory bandwidth. Reuters reported that next year’s deployment could potentially exceed one million chips; Microsoft’s official announcement, however, committed to a multi-gigawatt rollout rather than a specific chip count. Microsoft announcement, Reuters report

What stands out to me is that Microsoft is emphasizing throughput per unit of power, not just raw compute. It says Maia 300 is expected to process 40% more tokens per megawatt than leading GPU systems, and up to seven times the rack-scale token throughput of Maia 200 in a full 96-accelerator configuration. Those are Microsoft’s comparisons, not results independently verified under neutral conditions. For a real service, I care about how many paid requests can be handled reliably for the same energy and cost.

The chiplet design and the effort to co-design memory, networking, and software all point in that direction. Developers using Azure APIs may benefit from better cost or throughput without changing application code. Teams operating their own infrastructure will have to weigh accelerator-specific tuning against inference layers that can run across multiple chips. I don’t read this as Microsoft replacing Nvidia overnight. But when the largest cloud buyers also grow their own alternatives, competition expands beyond performance to power efficiency and supply reliability.

2. Meta puts Muse Wonder, an agent with longer-term memory, on paid paths

Meta introduced Muse Wonder and Muse Wonder Full on October 12. The company describes them as models for personal and professional agents. Wonder is available through the Meta Model API and Muse Code; Wonder Full is available to Muse Code subscribers. “Active Memory” carries relevant information across conversations, while “Active Focus” is designed to keep working when a task stalls, try other approaches, and recover from errors. Meta’s announcement, Reuters report

Reuters reported that Meta shares rose about 4% during afternoon trading on launch day. I would treat that as a market reaction that may reflect hopes for monetizing Meta’s AI investment, not proof that the model announcement alone caused the move.

I would not summarize this as “Meta has completely abandoned open models.” The new products have API and subscription routes, but Meta did not say in its launch post whether Wonder-family weights would be released. Axios also noted that the company had not clarified that point. What is confirmed is a new attempt to sell access to frontier-level models—not the end of Meta’s entire open-weight strategy. Axios report

Longer-term memory could make an agent more useful, but I think control over that memory matters as much as model quality. Users need to know where a remembered detail came from, how stale information can be corrected, what happens when they ask for deletion, and how memory affects actions taken through external tools. Meta says a fuller safety evaluation will follow. Until independent evaluations and memory controls are clear, I would not assume that a model remembering more automatically means personal information is safe.

3. When oil and the dollar rise together, AI operating costs feel it too

On October 12, Brent crude futures rose 2.3% to $107.17 a barrel, while the dollar index gained 0.37% to 102.58, its highest level since April 2025. That day, Cleveland Fed President Beth Hammack said inflation pressures were broadening beyond energy and influencing companies’ price-setting plans. She argued that, if needed, it would be preferable to move policy toward a restrictive stance relatively quickly, while leaving her October meeting decision open. Reuters market report, Cleveland Fed speech

I read these numbers as a cost-transmission chain, not as isolated headlines. When oil and the dollar both become more expensive, Korean companies can face higher energy-import bills as well as higher cloud, model-API, and semiconductor-equipment costs paid in dollars. If broader inflation keeps rates elevated, token prices alone won’t tell us whether an AI service is economical. Exchange rates, power, infrastructure depreciation, and actual paid usage all belong in the same calculation.

That does not mean AI investment is about to stop. It does mean that when a team adds a model, I want it to ask not only “Is it smarter?” but also “What does each request cost, and are there enough paying users to recover that cost?” Market prices can move quickly, so I would not infer a long-term trend from one day’s oil and currency moves.

4. India’s RBI changes the route through which oil companies buy dollars

India’s central bank announced the measure on October 10. From October 12, three state-run oil companies began obtaining their daily dollar requirements through a separate facility routed via designated banks. Reuters reported that the rupee rebounded about 1%, from Friday’s close of 96.73 per dollar to 95.77 on Monday, its largest daily gain since April. Reuters estimated that the facility could redirect $10–12 billion in monthly dollar demand away from the spot market. Local reporting has separately put the three companies’ daily dollar needs at about $200 million, so the estimates may not use the same scope. The Reuters report also said foreign investors had withdrawn more than $30 billion from Indian equities in 2026 and reserves had fallen from about $785 billion in September to $734 billion on October 2. Reuters report, local estimate of oil-company demand

The measure changes where the dollar demand is met; it does not erase the oil bill. The imports still need to be paid for, and when the central bank supplies dollars, reserve management and future liquidity matter too. Other FX measures and possible official market intervention may also have contributed to the rupee’s rebound, so I would not attribute a one-day move to a single policy.

My practical takeaway is that recording the exchange rate alone may not be enough. A company’s actual cost can depend on where it obtains dollars, the maturity, and the hedge terms. We still need to see whether the RBI’s arrangement lasts and whether reserves and capital flows stabilize. I’ll record this as an example of a central bank changing the plumbing of FX markets in response to an energy-import shock, not as a policy every country should copy.

What I’m taking away

Together, these developments make me ask again how AI becomes economical. Microsoft wants to scale its own chips to increase inference per unit of power; Meta has launched agents that can carry memory across sessions. At the same time, rising oil and a stronger dollar can change the cost of running data centers and APIs, while India’s case shows that central banks may alter the route through which FX demand reaches the market.

I don’t want to judge AI competition by model performance alone. I also want to know how efficiently the chip works in production, whether an agent’s memory is under the user’s control, and whether a service turns its costs into revenue. Maia’s efficiency figures remain company claims, and Muse Wonder’s weight-release plans and memory controls need further clarification. I found no new crypto or blockchain policy change or structural event today that clears the bar for a separate item, so I’m not repeating price moves.

This is my personal record of technology and market developments, not a recommendation to buy or sell any asset. I distinguish confirmed facts from company claims and market estimates. Each reader is responsible for their own decisions.