2026.10.05 | Substack | Weekly | Weekly: Gemini 4, MU Visibility & Optics

First of all, we are honored to be at Tomorrow X Summit in Austin on November 17–18 as a sponsor and partner, with more than ten team members attending and a session on how companies can build their own context layers for AI workflows. The FUNDA community can get 20% off passes through the registration link in our announcement. Join us and do let us know if you are visiting!

FUNDA

@FundaAI

We’re proud to be a sponsor and partner of the Tomorrow X Summit, taking place in Austin on November 17–18, 2026. Since launching in November 2024, FUNDA has been committed to producing and supporting the best independent research on Substack. We share Gavin’s view that …

2:53 PM · Sep 30, 2026 · 6.78K Views

1 Repost · 8 Likes

Before that, a couple of our team members will host a new Happy Hour in San Jose, CA on Oct 13 during the OCP Global Summit. If you are around, please RSVP and we will reach out once your registration is approved and confirmed. Look forward to seeing you there as well.

Coming back to research, Google’s return to frontier competition gives us another reason to stay constructive on compute demand. We argued in August that Google would keep investing despite the management departures and delayed model release; Gemini 4 now supports that view. Our work on open-weight models points in the same direction: as access to full reasoning traces tightens, developers will need to spend more on their own reinforcement learning, synthetic data and training environments. The recent low-cost catch-up looks difficult to sustain, while several well-funded labs remain in the race.

Micron puts longer-term demand into a more tangible frame. More than 75% of its CY27 output is already committed through SCAs and advance purchase orders, and customer discussions are shifting toward CY28 allocation. That supports our view of a higher, longer memory-price plateau. The headline gross-margin guide looked softer than the buy-side bar, but our adjustment for incentive compensation and new-fab costs puts it above expectations. We continue to focus on how long the supply shortage lasts.

Optics stocks also responded to the prospect of a more workable FCC framework, particularly Lumentum and Coherent. There is no public draft yet: the reported 65% U.S.-company BOM threshold remains a scenario under discussion. If adopted, it could support U.S. component suppliers while allowing leading Chinese module makers to keep serving customers. Our ECOC conversations separately reinforced the NPO ramp and the search for lower-power optical links; in power generation, FTAI’s first Mod-1 deliveries and customer acceptance in Q4 are the next concrete tests of another route to faster data-center energization.

This Week’s Reports

Gemini 4 and open-weight models: frontier competition continues to require more compute. Google is back in contention, while tighter access to reasoning traces makes cheap distillation harder and increases the need for developers’ own training investment.

Research|LLM: Gemini 4 Brings Google Back into Frontier Competition as Investment Accelerates

Oct 1

Google announced Gemini 4 Argon on September 30, roughly ten months after Gemini 3. On Google’s evaluations, Argon leads or ties GPT-6 Astra and Claude Opus 5.5 on 14 of 19 benchmarks. Artificial Analysis scores it at 53, level with Astra and below Opus 5.5 at 58 and Sonnet 5.5 at 56.

Read full story

Deep|LLM: Open-Weight Models’ Low-Cost Catch-Up Is Unlikely to Last

Sep 29

A succession of strong open-weight models were released this summer. Zhipu launched GLM-5.2 on June 22, followed by GLM-5.3 on August 14, a release focused solely on scaling post-training. On Zhipu’s own evaluations, GLM-5.3 matched or outperformed Fable 5 and GPT-5.6 Sol on several agentic coding tasks. Moonshot AI launched Kimi K3 on July 17. It scored 57 on the Artificial Analysis Intelligence Index, behind only Fable 5 and GPT-5.6 Sol. DeepSeek’s August 13 update to V4 Pro scored 87.9 on Terminal-Bench 2.1, nearly matching Fable 5’s 88 at roughly one-fiftieth of its cost per task. DeepSeek followed with V4.1 Flash in September. Xiaomi released MiMo-V2.6 with open weights on September 22; its Pro variant now leads the open-weight rankings on the Artificial Analysis Intelligence Index.

Read full story

MU: demand visibility now reaches 2028. More than 75% of CY27 output is committed, and management expects CY27–28 supply-demand conditions to be tighter than in 2026, supporting the longer price plateau we outlined before the results.

Review|MU FY26Q4: Visibility Extends to 2028, Supply Tighter Than in 2026

Oct 1

Micron expects memory supply-demand conditions in CY27 and CY28 to be tighter than in 2026. Customer orders and AI compute demand continue to rise. Even after accounting for planned cleanroom expansions across the industry, Micron sees no clear timeline for supply and demand to return to balance. Management expects prices to keep rising through FY27, though at a more moderate quarter-over-quarter pace. This is consistent with our prior view: memory prices will continue to rise, with the pace constrained by LTAs and customer affordability.

Read full story

Optics: NPO is moving toward products, while the FCC outcome could be better than feared. Our ECOC fieldwork tracks the architectures and supply constraints behind the ramp, and our policy column examines a reported U.S.-component threshold that remains under discussion rather than an adopted rule.

Research|Optics: ECOC 2026 Takeaways, NPO Gains Momentum as Slow and Wide Takes Shape

Sep 29

We attended ECOC 2026 in Málaga last week. The debate between slow and wide and fast and narrow was a recurring topic in presentations and plenary sessions. Companies presented microLED and micro-VCSEL designs and discussed their performance, manufacturing challenges and production timelines. NPO was also in focus, with live NPO demos on the show floor and more discussion of how and when it ramps.

Read full story

Column|Optics: FCC Transceiver Rules Might Be Better Than Feared, and Better for Top Chinese Players Than You'd Think

Oct 2

Investors have been discussing yesterday’s strength in optics stocks, particularly Lumentum and Coherent. The gains have been attributed to a widely discussed note reporting a senior legal team’s view of how the FCC may frame restrictions on Chinese-built optical transceivers.

Read full story

FTAI: aviation aftermarket capabilities meet the demand for earlier power. We initiate with a focus on Q4 Mod-1 delivery and acceptance, modeling about 80 deliveries in 2027 while assessing how aftermarket share gains can offset margin pressure.

Deep|FTAI: Aviation Aftermarket Meets AI Power

Sep 30

Executive Summary

Read full story

Meta Muse: identifying the software behind completed tasks. Our review maps 25 products from 24 listed companies across communications, productivity, travel and commerce, distinguishing tested workflows from identified connections and treating additional revenue as unproven.

Deep|Meta Muse: The Software Companies Behind Its Actions (Part 1: Public Companies)

Sep 29

We went under the hood of Muse to see which software services it calls as it carries out a user’s request. We followed the steps from finding information and accessing accounts to making calls, managing bookings and returning a result, then linked those steps to the companies behind the software. This first part covers 25 products from 24 listed companies. It shows where Muse could bring more activity to existing communication, productivity, travel, commerce and connected-device services—and where we can currently confirm only that a connection exists.

Read full story

Premium Report Snapshot

A portion of our research is reserved for Premium subscribers and is not distributed via Substack. Below is a snapshot of what Premium subscribers received this week beyond the Substack feed.

Deep | LLM: Computer Use and the Scope for White-Collar Automation

Subscribed

We estimate that computer use could eventually add about $1tn of annual revenue for AI providers, with automatable work equivalent to roughly 180mn full-time jobs, or 18% of global white-collar employment. These are estimates of work hours, not a forecast of layoffs; early adoption is more likely to reallocate tasks. Reliability, compliance and the Windows experience remain constraints, with administrative and operational workflows the most immediate opportunities.

Research | Muse and Consumer Cybersecurity: Mapping the Security Value Chain

Meta’s Muse shifts consumer identity from “person + devices” to “person + personal agent + sub-agents,” but Meta’s Sentinel captures much of the internal runtime/security layer, leaving third-party upside in external authorization, governance and recovery. We identify that OKTA, SAIL and PANW/CyberArk are exposed mainly to Step 4 Authorization and Step 9 audit/revocation as agents move from read to act to transact. Directionally positive but early; monetization depends on agent volume and IAM extension into consumer-agent permissions.

Deep | Meta Muse: The Software Companies Behind Its Actions, Part 2 — Private Companies

We identify connections to 21 private software companies across finance, payments, travel, productivity, development, health and device control; ten have documented workflows and eleven are confirmed connections only. Growing Muse usage could bring more API calls and transaction activity, but a connection alone does not establish a commercial agreement, usage frequency or incremental revenue.

Research | Servers: Power and Cooling Content Keeps Rising Through 2028

Our supply-chain work points to higher power and cooling content across NVIDIA VR300, Google TPU v9 and AWS Trainium4, supporting our constructive view of Delta and Infineon. We expect VR300 shipments in 27Q3, with double-sided cold-plate ASPs around twice current levels, while TPU v9 could carry roughly $400–500 of VPD content per chip. Trainium4 may enter trial production after 27Q3 with VPD and full liquid cooling, with Infineon returning to its supply chain.

Research | Space: Suncatcher Prototype Reaches Orbit

Our note reports that Google’s first Suncatcher prototype launched on October 1, carrying four Trillium TPUs for orbital testing. Radiation and thermal-control data over the next few weeks, including cooling intervals after roughly fifteen minutes of operation, are the immediate tests, followed by a planned two-satellite laser link in 2027. Launch demand and potential Planet orders offer nearer-term business opportunities, while profitable commercial orbital computing remains a much longer-term prospect, more likely in the early 2040s.

Weekly Expert Interviews Summary

A snapshot of the expert interviews we conducted during the past week is below; full transcripts and takeaways are available on the FUNDA platform.

Astra: Computer-Use Coverage and Deployment Economics

Engineer, AI research and deployment lab

The expert estimates coverage of about 70% of the top hundred enterprise applications, but only 30%–35% achieving task success above 90%, while a tested two-tier architecture cut token use from roughly 140k–150k to just over 30k.

Anthropic: Model Economics and the Computer-Use Opportunity (AMZN, GOOGL, NVDA)

Executive, Top AI Research Firm

The expert sees end-to-end agents expanding the opportunity beyond coding and Astra slightly ahead in interface-heavy tasks, while cautioning that capability does not guarantee revenue and some compute and lifetime-revenue estimates are inconsistent.

Anthropic: Model Tiers and Continued Frontier Investment (AMZN)

Executive, Global Leader in AI Systems Provider

The expert describes mid-tier models as the main monetization engine, with frontier models supporting retention, faster research and future lower-cost offerings despite weaker standalone economics.

Nuvei: AI Spending Under Compliance Constraints

Executive, Global Leader in Payment Solutions Provider

The expert estimates monthly AI spend rose from zero to $200k in six months through Copilot subscriptions and expects the next $100k of monthly spending to come from APIs, with adoption unmeasured and compliance, transaction risk and data readiness still limiting deployment.

FTAI: Mod-1 Production Ramp and Reliability (FTAI)

Specialist, Leading Aerospace Manufacturer

The expert considers annual output of one hundred units operationally feasible, with component availability and resource allocation the main constraints, while establishing a proven reliability record could take about three years.

Robotics: Data Bottlenecks and Training Economics (CRWV, NBIS)

Executive, Leading AI Solution Provider

The expert identifies real-world data collection and labeling as the main bottlenecks and expects no vendor to ship more than ten thousand humanoids in 2027, with adaptation to general environments still three to five years away.

Nittobo: Specialty Glass-Fabric Shortages and Capacity Expansion

Executive, Top chemical manufacturer

The expert expects low-dielectric and low-thermal-expansion glass shortages to persist because weaving and finishing capacity takes time to expand, with Nittobo’s product mix and Taiwan Glass qualifications shaping the supply outlook.