Key Takeaway
AI is shifting from a research tool into an economic operating layer. Agents are beginning to trade, transact, and coordinate through programmable payment rails, while falling model prices sharpen the divide between durable growth and an expensive narrative. Human approval, infrastructure risk, valuation, capital discipline, leverage, and risk still shape the investment case.
From Assistance to Agentic Finance
AI-powered trading is progressing from generating ideas toward monitoring earnings, harvesting tax losses, rebalancing portfolios, and executing approved workflows.
The practical model remains simple: AI informs, and the human decides. Obi reviewed every recommendation before trading. Ivy cannot place trades. Public plans to require users to approve workflows.
That restraint matters. One automated experiment involving roughly $200 was losing money consistently. Instructions such as “grow my portfolio aggressively” also create a basic problem: an agent must translate an ambiguous human goal into precise risk limits.
For now, the opportunity is better research and execution support, not unquestioned autonomy.
The agent thesis is also moving into commerce. Virtuals Protocol has facilitated about half a billion dollars of agent transactions and $2.5 million in agent profits, primarily in trading.
Its infrastructure includes smart wallets, ERC8183 escrow, and on-chain reputation. The goal is to enable programmable payments between agents, with spending rules, refunds, evaluators, and fee splitting built directly into transactions.
The constraints remain unresolved. Service completion is only about 80%-plus. Wallets face hacking and prompt-injection risks. Agent commerce has not reached product-market fit.
Virtuals’ strongest business has instead been agent assets, including about $15 billion in trading volume. Its base-pair structure has locked about 58% of the virtual token’s circulating supply in liquidity pools.
Agents are becoming more capable of acting, paying, and coordinating, but human approval and infrastructure risk still define the practical limits.
Falling AI Costs Meet Company-Specific Value
Sam Altman announced major price cuts for GPT-5.6 Luna and Terra. Luna fell 80% to $0.20 per million input tokens and $1.20 per million output tokens. That was down from $1 and $6, characterized by Zephyr as a 5x price cut and the opening of a price war.
GPT-5.6 Terra fell 20% to $2 per million input tokens and $12 per million output tokens. GPT-5.6 Sol gained an API Fast mode offering up to 2.5 times the speed for 2 times the price.
Lower inference costs could expand agent adoption, but company-level returns still depend on valuation and capital discipline.
Meta trades at 14 times next-twelve-month earnings, compared with 19 times for the S&P 500. Its projected five-year growth is 20%, versus 10% for the index.
Apple’s Q3 revenue was $109.4 billion, compared with $108.9 billion expected. Earnings per share were $2.02, versus $1.89 expected. iPhone revenue reached $54.3 billion, compared with $53.6 billion expected. Mac revenue was $10.4 billion, versus $8.6 billion. China missed at $18.8 billion, compared with $19.6 billion.
Apple increased research and development spending by 32% to $11.7 billion, potentially signaling a larger AI push.
Amazon reported Q2 revenue of $200.6 billion, earnings per share of $5.75, AWS revenue of $42.2 billion, and operating income of $27.5 billion. All four figures exceeded the cited estimates.
The tension is in Q3 guidance. Amazon guided to $200 billion of revenue, compared with $204 billion expected, and $24 billion of operating income, versus $25 billion expected.
Google combines a price-to-earnings ratio of 16 with 24% revenue growth and 82% cloud growth. But annual capital expenditures have risen from roughly $25 billion to $30 billion toward $200 billion, with another significant increase expected in 2027.
The broad AI story may be getting cheaper, but investment cases remain company-specific. Growth, valuation, spending, and guidance must be considered together.
Viewing the Pullback Through Valuation and Risk
Tom Lee attributes the AI sell-off to forced deleveraging across memory, semiconductors, and Korea-linked exposure.
Margin debt reached between $1.4 trillion and $1.44 trillion. Micron trades at 4.8 times forward earnings, a 1.7-turn discount to its 10-year average.
Lee believes the S&P 500 is near the end of the sell-off. He also notes that a previous DeMark nine bottom preceded a 21% rally.
The bookmarked dip-buying list includes Nvidia, Robinhood, Marvell, Nebius, Ondas, Amazon, SoFi, SPCX, Zeta, and Nokia.
But a list of companies is not an investment framework. The essential filter is the choice between passive exposure and active value investing, based on personal goals, valuation risk, and the ability to tolerate a 50% decline.
AI investment also carries geopolitical and ethical exposure. Classified Pentagon work has involved SpaceX, OpenAI, Google, Nvidia, Reflection, Microsoft, Oracle, and AWS. Alongside that work is a requested $54 billion for autonomous weapons.
Paramount’s Warner Brothers bid could leave the combined company with approximately $79 billion of debt, $4.5 billion in annual interest expense, and leverage of 6.6 times EBITDA. The financing depends heavily on the Ellison family’s $47 billion commitment. At the same time, 12 state attorneys general have sued to block the transaction.
Bottom Line
AI is becoming cheaper, more agentic, and more deeply connected to economic activity. The quality of an investment still depends on valuation, capital discipline, leverage, and risk.
Durable growth and an expensive narrative can look similar from a distance. The work is separating them.