Cheaper AI. Who earns the next dollar?

Wednesday brief · September 30

• AI economics: cheaper model usage can widen adoption, but more activity must eventually pay the infrastructure bill. Today we separate product announcements from earned revenue.
• Watchlist focus: Meta targets small businesses, Microsoft connects enterprise data, and Amazon adds an agent route. Distribution matters; profitable use matters more. [2,4,5]
• Wheel desk: three visual studies—Microsoft, Meta and Oracle—compare ownership economics, trend and cash exposure. Microsoft and Meta remain watch-only; Oracle is a pass; prior-session prices are not executable offers.

The lead · cheaper agents need paying customers

OpenAI’s September 29 DevDay announcements included GPT-6.1 Sol. Standard API rates are $2 per million input tokens, $0.10 for cached input and $10 for output. Lower unit costs can make more business tasks worth attempting, but the invoice depends on how an agent actually works. [1,3]

Our investment read: the next phase is a contest over useful work completed per dollar. A cheap request that needs repeated retries or extensive human correction can cost more than a pricier request that succeeds. Investors should ask about total task cost, completion quality and customer retention together. A benchmark victory alone cannot answer those questions.

That creates different opportunities across Eddie’s watchlist. Amazon and Microsoft sell infrastructure and enterprise access; Meta brings distribution and an advertising-funded business. Model efficiency may expand usage while putting pressure on price per task. Suppliers can grow volume without retaining the same margin on every unit.

The counterargument deserves equal weight: demand might grow quickly enough to offset cheaper inference. We would look for repeat paid workloads, measurable customer savings and cash generation after investment. Product launches establish new routes to market; they do not establish how much customers will spend or how profitable that spending will be.

Eddie’s watchlist · four signals to follow

• META — merchant workflows: Muse Small Business adds connections including Shopify, Canva and QuickBooks. Meta says publishing, sending and spending require approval. Our question is whether merchants keep using it after onboarding; tool availability is not evidence of sales uplift. [4]
• MSFT — enterprise context: Microsoft says Fabric IQ is generally available in Copilot Chat and Cowork. Other capabilities remain previews or forthcoming. Governed business data could improve useful answers, but integration work and customer trust can slow monetization. [5]
• HOOD — activity, not a full month: Robinhood reported 234 million options contracts for September 1–24. This preliminary, unaudited window should not be extrapolated mechanically into quarterly earnings. Higher participation can support revenue while increasing sensitivity to market conditions. [9]
• MU — results ahead: tonight’s earnings call is a fresh checkpoint for memory economics. Focus on capacity, pricing and customer commitments together. We exclude Micron from the wheel desk before this scheduled event rather than treating option premium as free income. [11]

AI buildout & adoption · who captures the value?

Amazon gains another distribution route. AWS describes Bedrock Managed Agents powered by OpenAI as a preview, developed with OpenAI; runtime and inference stay within AWS. OpenAI separately confirms the collaboration. Our interpretation: enterprise identity, permissions and existing cloud relationships may matter as much as the model choice. Availability is not evidence of broad production adoption or incremental AWS profit. [2,3]

Oracle illustrates the financing test. Its August-quarter infrastructure-cloud revenue reached $7.4 billion, up 121%, while the same quarter’s free cash flow was negative $5.4 billion. Operating cash of $23.1 billion did not cover $28.5 billion of capital spending. Growth does not remove the need to fund construction. Follow cash conversion and financing alongside booked demand; a large contract balance is not cash already collected. [8]

One comparison · caching helps one part of the bill

Text alternative: standard Sol input costs $2, cached input $0.10 and output $10 per million tokens. Cached input is 95% cheaper than standard input; that is not a 95% saving on an entire workflow. Output, tools, storage, retries and review still matter. A million input tokens and a million output tokens are different billing categories, not equal business outcomes. [1]

Macro lens: today’s GDP and inflation releases could alter rate expectations. A hotter inflation print may pressure long-duration valuations; softer inflation may help financing conditions while leaving demand questions unresolved. These are scenarios, not forecasts. Correlated technology exposure can overwhelm the apparent diversification of three different tickers. [10]

3 Wheel Trade Ideas

WATCH-ONLY / PASS · October 16 puts, 16 calendar days. Stocks: September 29 completed session. Cboe retrieval batch: September 30, 06:55 Toronto; this is retrieval time, not quote time. Last trades are September 29; refreshed Greeks may reflect premarket inputs. Exact bid/ask time, timezone, delay, open-interest date and IV rank/percentile remain unconfirmed. Options are closed at 08:00; reprice after the 09:30 open. No fill assumed.

One standard 100-share contract per study; verify deliverables. Cash = strike × 100; credit = prior-session bid × 100; breakeven = strike − bid; maximum loss = cash − credit if shares reach zero. Wilder RSI14 and SMA20/50/200 use Yahoo adjusted closes; relative volume uses the preceding 20 full sessions. All amounts USD. Delta is not assignment probability. [12,17–22]

Valuation ruler: breakeven divided by four times the latest quarterly GAAP diluted EPS gives a simple earnings proxy. It assumes the quarter repeats, ignores seasonality and can be distorted by unusual charges. It is neither forward P/E nor fair value. Compare financing and business quality before comparing the multiples.

MSFT · Microsoft

Watch-only · stronger trend · Prior-session study: the $480 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. Reference arithmetic; not a current offer.

• Why own the shares: Enterprise software and Azure support the ownership case, with rate-sensitive valuation risk. June-quarter revenue was $90.0 billion and GAAP diluted EPS $4.81. GAAP earnings included a $3.2 billion Anthropic investment gain. Reference breakeven is 24.82× annualized quarterly EPS. Repeat cloud demand must cover infrastructure costs; this proxy is not a bargain verdict. [6]
• Trend and levels: Close $508.96; RSI 58.5; SMA20/50/200 $499.91/$479.88/$430.88; relative volume 1.00×. Above all three averages, not oversold. [17]
• Contract quality: Bid/ask $2.47/$2.78; spread $0.31; delta −0.1588; IV 28.39%; volume 343/OI 8,185. Last trade September 29, 15:58:50, provider clock. [18]

META · Meta Platforms

Watch-only · do not chase momentum · Prior-session study: the $680 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. Reference arithmetic; not a current offer.

• Why own the shares: Advertising finances the AI opportunity, but the economy and marketing budgets remain risks. June-quarter revenue rose 28% to $60.8 billion while operating income fell 8%; legal and severance charges affected results. Breakeven is 27.29× annualized $6.18 quarterly GAAP EPS; charges weaken comparability. [7]
• Trend and levels: Close $738.79; RSI 64.6; SMA20/50/200 $675.73/$618.59/$626.18; relative volume 1.00×. Strong momentum, not an oversold entry. [19]
• Contract quality: Bid/ask $5.30/$5.40; spread $0.10; delta −0.1565; IV 40.36%; volume 326/OI 1,800. Last trade September 29, 15:59:19, provider clock. [20]

ORCL · Oracle

Pass · repair the trend first · Prior-session study: the $125 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. Reference arithmetic; not a current offer.

• Why own the shares: AI infrastructure demand is strong, but funding and dilution matter when rates stay restrictive. August-quarter revenue was $19.3 billion; GAAP diluted EPS $1.56. Breakeven is 19.83× annualized quarterly EPS. The smaller multiple and cash requirement do not eliminate negative free cash flow or financing risk. [8]
• Trend and levels: Close $137.79; RSI 43.1; SMA20/50/200 $147.16/$142.70/$162.96; relative volume 1.33×. Below all three averages; RSI is not oversold. [21]
• Contract quality: Bid/ask $1.29/$1.55; spread $0.26; delta −0.1778; IV 51.32%; volume 3,560/OI 10,933. Last trade September 29, 15:54:49, provider clock. [22]

Today’s catalysts · Toronto time

• September 30, 08:30: BEA third estimate of Q2 GDP. Separate revisions to past growth from evidence about demand now. [10]
• September 30, 08:30: August personal income/outlays, including PCE inflation. No released figure is assumed in this premarket edition. [10]
• September 30, 16:30: Micron’s earnings call. Examine memory demand and capacity commitments before extrapolating AI spending. [11]

From YieldCove · ownership before premium

A later covered call can cap upside without repairing a large stock loss. Before any wheel, write down why you would own 100 shares through a difficult quarter. Compare that commitment with your existing exposures, not only the premium. Explore YieldCove to organize the assumptions; cash is an available position too.

AI-assisted editorial by Codex. Educational information, not personalized investment advice. Early assignment is possible. Scenarios exclude fees and taxes; no trades are executed.