SPECULATION — scenarios, not announcements · Iggy’s forecasts, Oct 2026 → Oct 2028. Not company announcements.
SPECULATION — scenarios, not announcements
Iggy's directional forecasts for Oct 2026 → Oct 2028. 'Compute signal' is public infrastructure or product evidence, not a product-specific allocation. Confidence is about direction, not timing. Nothing on this page is a company announcement or a measured compute allocation.
Product trajectories, conditioned on compute Speculation
Compute helps most when it turns into reliable inference capacity, better serving economics and more useful agent work — not merely a bigger training run.
Categories are editorial groupings for browsing, not company classifications.
Scenario · not an announcement
Consumer direction confidence
Now · public signal
Muse Wonder / Wonder Full launched Oct 2026; Meta says Wonder will soon power Muse. Prometheus ~1 GW; Hyperion up to 5 GW (longer-dated).
If capacity lands · to Oct 2028 (speculative)
More capable persistent memory, multimodal and professional tasks; more background and parallel work. Smaller models may broaden access while frontier stays premium-first.
Confidence & watchpoints
Medium-high on capability investment; medium on adoption. Watch trust, subscriptions, share of Meta's fleet serving Muse.
Bottom line: capacity makes ambitious agent loops affordable and available; it doesn’t guarantee useful autonomy. Expect competition to shift from “who has the largest model?” to “who completes more trusted work per unit of inference spend?”
SPECULATIVE PRODUCT CONCEPT — not a Google product or roadmap item
Gemini Relay (working title)
“Tell Relay the outcome. It plans, does the reversible work, and asks before it commits.”
A single task inbox across Gemini, Gmail, Calendar, Drive, Docs, Sheets, Chrome and Android. Hand off a task from the screen you're on, see a short plan, watch the agent work in a cloud workspace. Each task has a persistent brief, a visible activity log, and one-tap stop/undo.
Why Google could be strong: Workspace and Android/Chrome distribution, its own TPUs for a tiered inference router, Search and Workspace context. Caveat: Alphabet's compute is shared with Cloud customers and other services; Relay would have to earn capacity through efficient serving.
MVP: explicit read/draft/act scopes for Gmail, Calendar, Drive; plain-language outcome; plan with deadline and bounded budget; sandboxed drafting; confirmation before sending, purchasing, deleting or sharing; concise receipt plus undo.
Differentiator: a fast, cheap model for routine steps, escalating only for hard ones; show the compute/time budget; measure accepted work per inference dollar.