AI meets the grid queue: costs, controls and three wheel decisions

MONDAY · SEPTEMBER 21, 2026 Before the U.S. open. Market studies use Friday, September 18; dollars are USD. Weekend developments and earlier operating evidence are dated separately.

The morning brief

• AI meets the grid queue. AWS is applying agents to utility studies; faster preparation is a narrower achievement than faster construction. [1]
• The operating bill matters. Google’s new storage controls and a power-flexibility alliance target costs beyond model inference. [2–3]
• Three ownership decisions. Alphabet is conditional; Oracle needs a recovery; Credo remains a pass. Every premium accompanies an obligation to own shares.

The lead · Can AI shorten its own power bottleneck?

AWS’s September 17 grid-planning program offers a concrete application of agents to an infrastructure bottleneck. AWS reports that collaborating utility Duke Energy reduced data preparation from two weeks to hours. Engineers retain final decisions, while existing simulation software performs the deterministic analysis. Agents coordinate the workflow. This is a company-reported improvement in preparation, not independent evidence that a complete interconnection or construction process now takes hours. [1,13]

Our interpretation: the commercial opportunity lies in making scarce engineering time more productive. Repeatable preparation, documented assumptions and reviewable results could let a utility evaluate more projects without removing human responsibility. Amazon may gain cloud and integration work; AI infrastructure customers may eventually benefit from more efficient studies. Neither revenue contribution nor additional connected megawatts is quantified in the announcement.

The counterargument is physical and institutional. A better study cannot manufacture transformers, acquire land or settle who funds network upgrades. Additional scenarios may also reveal constraints that postpone a project. Investors should therefore follow three milestones: adoption beyond the initial utility, measured end-to-end study times, and connections actually approved. Treat the first workflow result as evidence worth investigating, while withholding a blanket forecast for faster AI capacity growth.

Eddie’s watchlist radar

• GOOGL · Costs become visible. Google’s September 18 digest lists Storage Intelligence Advisor as generally available, flagging unusual activity and egress costs. Watch whether customers retain savings after usage expands. [3]
• ORCL · Backlog needs conversion. September 10 results put remaining performance obligations at $664 billion. The next test is funded delivery and collected cash; backlog is not current revenue. [6]
• CRDO · Growth has working-capital demands. The August 1 quarter’s filing shows $151.3 million of working-capital cash outflows. Monitor collections and inventory as connectivity demand grows. [8]

AI buildout & revolution

Flexibility before more steel. The September 16 alliance from Emerald AI, Google and NVIDIA proposes measurable power-response commitments to support grid connections. Shifting workloads or using storage can help, but operating performance and utility acceptance must be demonstrated. The announcement does not establish cheaper electricity for every customer. [2]

The weekend signal. NVIDIA’s September 20 X post argues that rapid development can coexist with testing, monitoring and safeguards. That is a stated position, not proof of system safety. For investors, the useful follow-up is measurable reliability and the cost of oversight in paid deployments. [4]

A live grid stress test. A September 19 DOE notice describes a temporary order effective September 18–21 authorizing specified Duke resources and, under emergency conditions, backup generation amid Carolinas heat. It illustrates why operating flexibility matters. The order does not establish that AI caused the strain. [20]

The comparison · How much revenue goes back into equipment?

Oracle’s equipment spending exceeded quarterly revenue; that needs funding. Credo’s low ratio reflects a different model and excludes its $735.6 million acquisition cash payment. This is a capital-intensity comparison, not a ranking of returns or financial strength. [5–6,8]

3 Wheel Trade Ideas

Read this first. With the Fed at 3.75–4.00%, reserved cash has an opportunity cost. These are 25-day studies to October 16, not orders. Friday option data require repricing at today’s 09:30 Toronto open. Exact bid/ask timestamp, provider timezone, delay, OI date and IV rank are unavailable; displayed last trades are not quote timestamps. [9,12]

Daily Wilder RSI14 and SMA20/50/200 use adjusted completed closes; relative volume compares Friday with the previous 20 sessions. None is oversold. Next earnings dates were not confirmed in accessible IR calendars. Verify events and standard 100-share deliverables with the broker. Delta is not an assignment probability.

GOOGL · Alphabet

Conditional · require support to hold · The setup: sell the $325 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. A discounted entry still needs a valuation limit.

• Why own the shares: Diversified demand supports ownership, but rates raise the valuation hurdle. June quarter: 34.0% operating margin, −$5.86bn FCF; $242.47bn cash/securities versus $98.17bn long-term debt. Entry is 8.2× annualized quarterly sales per period-end share; not a forecast. [5]
• Trend and levels: Close $349.54; RSI 55.1; SMA20/50/200 $341.68/$345.40/$337.09; volume 1.98×. Above all three averages; seek stability around $337–345 rather than chase. [14]
• Contract quality: Bid/ask $2.83/$3.05; spread $0.22; delta −0.1780; IV 30.57%; volume 566/OI 5,886. Last trade September 18, 15:55:59, provider clock. [15]

ORCL · Oracle

Watch only · trend and funding first · The setup: sell the $135 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. Do not mistake a smaller cash ticket for lower business risk.

• Why own the shares: Cloud demand is strong, but financing matters at higher rates. Quarter ended August 31: 34.8% operating margin, −$5.40bn FCF; $37.08bn cash/securities against $125.34bn borrowings, excluding leases. Entry is 21.2× annualized quarterly GAAP EPS, not a forecast. [6]
• Trend and levels: Close $147.61; RSI 49.6; SMA20/50/200 $149.41/$141.00/$165.47; volume 1.35×. Below short/long averages; wait for a sustained $150 recovery while $141 holds. [16]
• Contract quality: Bid/ask $2.67/$2.80; spread $0.13; delta −0.2304; IV 47.35%; volume 923/OI 9,848. Last trade September 18, 15:59:37, provider clock. [17]

CRDO · Credo

Pass · require better price and trend · The setup: sell the $150 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. Higher IV compensates for uncertainty; it does not remove it.

• Why own the shares: AI connectivity offers growth, but customer spending is cyclical. August 1 quarter: 25.2% operating margin, $90.2m operating cash; $764.3m liquidity versus $284.2m total liabilities, no separate borrowing line. Entry costs 54.8× GAAP or 30.6× adjusted annualized quarterly EPS. Neither is a forecast. [7–8]
• Trend and levels: Close $175.89; RSI 43.1; SMA20/50/200 $188.85/$213.66/$175.83; volume 1.70×. The rebound barely holds the long average; this is not oversold. [18]
• Contract quality: Bid/ask $3.20/$3.40; spread $0.20; delta −0.1730; IV 68.02%; volume 319/OI 2,375; displayed bid depth four contracts. Last trade September 18, 15:59:46, provider clock. [19]

Next on the radar

• Today · 09:30 Toronto: reprice and inspect bid depth; Friday quotes cannot set Monday execution.
• October 9 · Oracle record date: confirm the exchange ex-dividend date separately before planning ownership. [6]
• October 14 · 08:30 Toronto: September U.S. CPI, two days before these options expire. [10]

From YieldCove

Use the wheel calculator to compare cash committed with downside loss. Covered calls after assignment can cap recovery; premium is not a substitute for a sound ownership case.

Photo: Googleplex, Mountain View, May 25, 2013. Austin McKinley / Wikimedia Commons, CC BY 3.0. Historical corporate context, not a data center; Wikimedia 1,280px rendition, no crop. [11]

AI-assisted research and translation, separately counter-verified by Codex. Independent Grok review is required before subscriber delivery.