AI’s next mile: networks, paying users and three wheel decisions

SUNDAY · SEPTEMBER 20, 2026 The week ahead. Market studies use September 18 completed-session data; all financial amounts are USD. This briefing combines fresh weekend announcements with explicitly dated context.

The morning brief

• Distribution broadens. Muse’s weekend release recap adds a Mac app and Canadian availability on iOS/web. These are company-reported launches, not evidence of paid retention. [2]
• The buildout spreads beyond GPUs. Optical capacity and memory-access research address different bottlenecks. Investors still need shipment and customer-economics evidence. [4–5]
• Three wheel decisions. Microsoft is conditional, Meta needs a calmer entry, and Marvell remains a pass at this valuation. None is technically oversold.

The lead · Who keeps the value from cheaper AI?

Meta’s September 15 subscription announcement puts an economic question beside the AI capability race. The company reported 15 million subscriptions and trials across Meta One plans; that combined figure is not 15 million paying customers. Its weekend Muse update adds distribution through Mac and Canada. More ways to try a product can widen the funnel without establishing its eventual revenue or profit. [1–2]

Our interpretation: the next investment test is the cost of serving a useful, repeatable task. Better networking and inference software may reduce that cost, but shareholders benefit only if customers keep using the service and the provider retains some savings. A discounted subscription, free trial or impressive demonstration is an input to the story, not the finished business model.

The counterargument is that cheaper AI can unlock entirely new demand. A lower price per task may support more total spending if agents become useful in additional workflows. That would help both platforms and infrastructure suppliers. Watch conversion from trial to paid use, sustained task completion and cash generation after equipment spending. For a wheel investor, the practical question is whether the proposed ownership price remains acceptable when adoption takes longer and costs fall less quickly than hoped.

Eddie’s watchlist radar

• META · Weekend distribution. Muse’s September 19 recap names Mac, Canada on iOS/web with Android still to come, Granola/Notion connectors and a developer platform. We treat these as rollout claims; next comes evidence of usage and retention, not extrapolated subscription revenue. [2]
• MSFT · Serving quality, not just speed. GAUSS, listed September 18, models time to the first token, spacing between tokens and sustainable request rates. It is a research simulator, not a disclosed Azure margin improvement. Next: validation on real workloads under customer latency requirements. [3]
• MRVL · Demand meets valuation. Revenue for the quarter ended August 1 was $2.739 billion, up 37%; data-center revenue grew 46%. These are reported results from August 27, not weekend sales. The October 6 investor day is the next dated opportunity to test capacity plans against financial expectations. [8]

AI buildout & revolution

Light needs manufacturing capacity. GlobalFoundries and Marvell announced a multiyear expansion of silicon-germanium capacity in Burlington, Vermont, on September 17. It supports pluggable and more tightly integrated optical connections. The release gives no dollar value or added wafer count. Our reading: network components deserve scrutiny alongside accelerators; an agreement still needs timely production and profitable orders. [4]

Reduce repeated memory traffic. Microsoft’s September 18 Capillary research uses shared memory across parts of Nvidia Hopper GPUs to avoid repeatedly fetching common context. This is a proposed kernel design; the page does not establish a fleet-wide savings percentage. The investment implication is a question: can software extract more useful work from installed hardware before operators buy another generation? [5]

One comparison · What remains after equipment spending?

The common formula exposes equipment intensity, not business quality. Meta’s official quarterly free cash flow is lower still because it also subtracts finance-lease principal. Marvell’s fabless model shifts manufacturing investment toward suppliers; its taller bar does not remove supply dependence. [7–8]

3 Wheel Trade Ideas

Weekend studies, not live orders. All expire October 16: 26 calendar days away. Assume standard 100-share puts; verify deliverables. Sunday Cboe downloads contain September 18 last trades. Exact bid/ask time, provider timezone, delay and open-interest date are unconfirmed. Reprice and revalidate Monday, September 21 at 09:30 Toronto. IV rank is unavailable; delta is not assignment probability. [15–21]

Technicals: daily Wilder RSI(14), 20/50/200-session SMA on adjusted closes; Friday volume divided by the prior 20-session average. RSI below 30 is a common oversold screen; none qualifies. The Fed’s 3.75–4.00% range raises the opportunity cost of reserved cash; October 14 CPI precedes expiry. [9–10,16,18,20]

MSFT · Microsoft

Conditional · wait for confirmation · The setup: sell the $470 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. Illustrative credit at Friday’s bid; no fill assumed.

• Why own the shares: Cloud/software distribution supports ownership; June-quarter cash after equipment was $19.64bn, with $76.84bn cash/short investments versus $40.29bn debt excluding leases. GAAP operating margin: 45.1%. Effective entry is 25.9× FY2026 GAAP EPS; investment gains mean earnings quality matters. [6]
• Trend and levels: Close $493.78; RSI 53.3; SMA20/50/200: $497.51/$464.20/$430.08; volume 1.97×. Above long averages but below SMA20: wait for a sustained $498 recovery. [16]
• Contract quality: Bid/ask $4.40/$4.60; spread $0.20; delta −0.2202; IV 25.39%; volume 1,594/OI 5,777. Last trade Sep 18 15:59:57, provider clock. [17]

META · Meta Platforms

Watch only · let the pullback settle · The setup: sell the $600 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. A put earns no dividend unless shares are owned in time.

• Why own the shares: Advertising funds AI experimentation, but capex limits cash flexibility. June-quarter GAAP margin 30.9%, official FCF $0.784bn; cash/securities $90.26bn versus long-term debt $83.66bn, excluding leases. Entry is 24.0× annualized quarterly GAAP EPS, not a forward or TTM multiple. [7]
• Trend and levels: Close $665.75; RSI 65.8; SMA20/50/200: $614.61/$607.38/$623.55; volume 1.53×. Friday’s drop did not create oversold conditions. Wait for stabilization near $624–$615 support. [18]
• Contract quality: Bid/ask $5.55/$5.80; spread $0.25; delta −0.1495; IV 38.52%; volume 1,433/OI 6,773. Last trade Sep 18 15:59:45, provider clock. [19]

MRVL · Marvell Technology

Pass · growth does not settle the price · The setup: sell the $210 put, expiring October 16, 2026. One contract means agreeing to buy 100 shares at that price if assigned. A higher premium accompanies materially higher volatility.

• Why own the shares: AI networking demand is attractive, but financing-sensitive customers can delay orders. Quarter ended August 1: GAAP margin 16.8%; operating cash less equipment $478.8m; cash $3.93bn versus long-term debt $4.96bn, excluding leases. Entry is 155.6× annualized quarterly GAAP EPS, or 54.6× non-GAAP: neither is a forecast. [8]
• Trend and levels: Close $244.25; RSI 58.0; SMA20/50/200: $227.48/$215.95/$158.36; volume 1.05×. Trend is positive, but shares sit well above intermediate support. No oversold entry. [20]
• Contract quality: Bid/ask $4.55/$4.80; spread $0.25; delta −0.1774; IV 66.45%; volume 299/OI 3,961; displayed bid size just 1. Last trade Sep 18 15:59:49. [21]

Next on the radar

• September 21 · 09:30 Toronto: reprice all contracts, confirm dividends/earnings and inspect bid depth after the open.
• October 6 · Marvell investor day: watch capacity economics and revenue assumptions; confirm the event time with IR. [8]
• October 14 · 08:30 Toronto: U.S. September CPI, before all three expirations. [10]

From YieldCove

Use the wheel calculator to compare reserved cash with a downside scenario. Owning an assigned stock remains the central obligation; selling calls afterward can cap a recovery.

Photo: Meta’s Menlo Park headquarters sign, May 12, 2022, Nokia621 / Wikimedia Commons, CC BY-SA 4.0. Historical context, not a Muse launch photograph; original unchanged. [14]

AI-assisted research and translation, counter-verified by Codex. Company statements and interpretations distinguished; independent Grok review required before subscriber delivery.