Excited to be part of @TomorrowXSummit in Austin on November 17 and 18, 2026! Great to see independent creators and the companies shaping AI coming together. Thanks to @GavinSBaker , @Atreidesmgmt , @rbiscardi and @iconnections_io for making it happen! FUNDA will also take the stage to discuss how institutions can build their own context layers to support AI-powered equity research workflows. More than 10 members of the FUNDA team will be in Austin. See you at Tomorrow X!

We had a great conversation with @vikramskr today! Vikram runs both semiexponent.com and semidoped.com, and we’re big fans of his work. We’re excited to collaborate with Vikram! If you’re also interested in partnering with us on our expert interview database, please get in touch!

Deep|LLM: Enterprise AI Adoption Survey (Vol. 5) - Cost Optimization and Application Expansion Proceed Together; Business Returns Shape Further Investment AI spending still rising: Across 10 enterprise interviews conducted September 15–20, most companies still expect AI budgets to increase. Growth is slowing where subscription seats are near full coverage, but API projects, agent rollouts and broader workflow deployment continue to lift spending. APIs become the next growth driver: A payment technology services company expects monthly AI spending to rise by approximately US$100,000 over the next approximately 6 months, entirely from API projects. A telecom operator is still expanding paid AI seats from below 10% of employees toward 30%–40%, with subscriptions expected to contribute approximately 80% of incremental spending. Cost optimization and usage growth coexist: Model switching, caching, deduplication and procurement negotiations are reducing unit costs, but total spending often keeps rising as applications scale. A digital marketing technology and services company cut model costs sharply but still sees spending grow approximately 2%–6% per month with customer and traffic growth. Returns shape future budgets: Enterprises increasingly require AI projects to show cost savings, revenue gains or measurable productivity improvements. A telecom operator replaced a marketing automation platform with an in-house application, avoiding annual license fees of approximately US$200,000–250,000, while a used-car marketplace reports positive returns on approximately 35%–40% of AI projects. Detailed Report fundaai.substack.com/p/deepl…

Preliminary thought on Oracle Project Jupiter Headline today & BE read-across: We don't think there's much incremental here vs what we already know. ORCL is essentially trying to give itself some contractual cushion in case the project runs into further delays, which seems like a pretty normal corporate risk management move to me. It doesn't mean Oracle is walking away from Jupiter. The gas infrastructure and permitting issues have been out there for a while, so the force majeure notice doesn't really change what we already knew about the project. What we'd be watching much more closely is how the New Mexico Supreme Court proceedings play out, when the air-permit process can get back on track, and whether the gas pipeline can be ready in time. Those are the things that could actually push out Bloom's commissioning and acceptance timeline. Project delays are always quite common for BE's deliveries. Also worth keeping in mind that Oracle's payment obligations to the developer and Bloom's equipment deliveries are two separate things, so we wouldn't automatically translate a potential delay in Oracle's data center COD into a delay in BE revenue. For now, we'd stay focused on the actual project milestones but are also awaiting more details and updates on this matter. Feel free to share your thoughts here as well.

Following our discussions about Muse last week and the release of our report on Monday, we received a lot of questions. 1. Can OpenAI and Gemini build a personal agent as good as Muse? We’ve spoken with quite a few researchers and teams building personal agents. Our main takeaway is that success depends less on the gap between models today and more on how seriously a company commits to the product: how many people it puts behind it, how much it spends, and how quickly it learns from users. In some ways, this feels like the early debate around Manus. The model matters, but so does all the work around it. You need a large team to understand what people actually want, improve individual use cases, and get the agent working reliably with other apps. One difference already stands out to us: Muse feels more like a consumer product, while the alternatives feel more like productivity tools. That distinction matters. Helping people with everyday needs is a different challenge from helping them get more work done. It’s also expensive. Beyond the product and integration work, there’s the compute bill. Based on what we’ve learned, running the service currently costs roughly $15–20 per DAU per month. There haven’t been many breakout consumer products in the US lately. But when Zuckerberg sees an opportunity like this, he tends to go all in. We think Meta has a better shot here than it did in VR or in the race to build the best AI models. Understanding users, refining use cases, and putting substantial resources behind a product are things Meta does well. 2. How do we think about Muse’s ROI? The biggest opportunity, in our view, is better purchase-intent data: a clearer understanding of what users are considering buying. In our best-case scenario, those additional signals could lift revenue by 25%. Using our estimate of roughly $50 in US revenue per DAU, that would mean another $12.50. Costs also have room to come down. We think better efficiency could eventually bring Muse’s running cost to around $10 per DAU per month. If Meta can get there, the advertising upside, together with other potential revenue streams, could more than cover the cost—even without a subscription fee. 3.Could Muse take ad share from Google? We think it could over time, but we don’t expect a meaningful impact over the next year. Muse gives Meta more purchase-intent data—a clearer picture of what users are actively considering buying. Over time, those signals could help Meta close the gap with Google in categories like travel, auto, and finance, and win some of the ad spending that currently goes to Google. But that will take time. Muse needs to build a regular user base, and Meta needs to turn those signals into better results for advertisers. We see this as a longer-term opportunity, not a near-term threat to Google’s ad business. 4. Can Tencent and Alibaba build a personal agent that is just as valuable? We don’t see the same incremental revenue opportunity in China, at least based on Muse’s current use cases. Many of those use cases address problems that are less common in China. Customer service is already efficient, for example, so people have less need for an agent that spends time making calls on their behalf. Tencent can also identify purchase intent much earlier through its existing apps and services. A personal agent may therefore add less information than it would for Meta. That doesn’t mean Tencent or Alibaba can’t build a useful product. But we wouldn’t assume the same revenue upside we see for Muse.

Tesla’s quick ramp of Optimus is positive for Intel Foundry. As mentioned in our recent Intel report, Tesla is engaging with Intel on its next-gen chip (likely AI6) to diversify supply chain risks, and we see a solid probability that Intel Foundry will secure a share of the manufacturing orders. Tesla’s Optimus Hits Snags in Hands, Suppliers as Scale-Up Begins — The Information

What developers use it for In week one, public discussion was mostly opinion; few people had actually used or tested Jev. Opinion and analysis are 41.8% of the 6,277 counted cases (Figure 1). The 2,153 hands-on cases are the better read on demand; the rest of this note uses that cut. Among hands-on cases, the largest uses are real-time control, agent control decisions and content classification. None reaches 20%. They are 18.8%, 16.1% and 15.7% of the 1,284 hands-on cases with an identifiable use (Figure 2). Table 1 shows a typical case for each.

Maybe a prick and not a puncture in equilibrium. type of curve that creates a generation of new startups "GPT-6 Luna is the first U.S. model since DeepSeek launched its V4 series to undercut DeepSeek’s cheapest model on both input and output list prices."

Deep|LLM: Jev Users Report 10× Faster and 54.5× Cheaper Than the Models They Replaced; Only 3.7% in Production Jev is a “decision model” from TypeSafe AI, released September 15, 2026 and opened to all users on September 20. It does not generate text. It answers questions with a fixed answer set: pick an option, score on a scale, or judge true/false, and attaches a confidence score. The launch quickly gathered industry interests, and some investors were asking whether it’s a significant negative to compute demand. As we addressed in our report earlier, we disagree with that concern and believes Jev is more of an interesting trial with limited impact on LLM. To analyze Jev further, we decided to have a deep dive into what Jev use cases are really about. This note covers 6,277 public discussions and use cases from the first 7 days; 2,153 are from people who actually used or tested it. -Demand sits on fast decisions with a fixed answer set. No single use clears 20%. Of the 1,284 cases with an identifiable use, the largest groups are real-time control in games, robots and simulations (18.8%), agent control decisions (16.1%) and content classification (15.7%). -Indie developers dominate the conversation; big-company engineers barely show up. Of the 2,140 authors whose role we could identify, 35.9% are indie developers, 23.4% are AI creators and KOLs, and just 2.7% are engineers at large companies. -Speed: 10× faster than the model it replaced or was tested against. Median user-reported speed-up is 10× (n=72): 10× vs frontier models, 5× vs small models. In the 16 cases with latency for both Jev and the prior system, Jev’s median is 300 ms vs 2,924 ms. The vendor’s 193.6× is a peak against the most expensive model. -Cost: 54.5× cheaper than the comparison model; the saving depends on what it replaced. Median user-reported cost multiple is 54.5× (n=56): 188× vs frontier models, 17× vs small models. The vendor’s own comparison with GPT-5.6 Terra is about 76×; the 444.6× in marketing is a peak against the most expensive model. -Accuracy: Jev and the systems it replaced each win some head-to-heads; gaps are small. In the 17 cases with accuracy for both, Jev is ahead in 10 and behind in 7; median gap is 1.6 percentage points. Of 241 cases that assessed accuracy, 83 rated Jev better and 65 worse. -Jev’s confidence scores miss by about 10 percentage points on average, and run clearly high on unfamiliar rating questions. Median user-measured ECE (expected calibration error: average gap between stated confidence and actual accuracy; 0 is perfect) is 0.097 (n=27). An independent test on unfamiliar tasks found 0.107 overall, but 0.325 on rating questions, where Jev was right only 44.7% of the time. -Developers put cheap small models next to Jev almost as often as the strongest large ones. Of the 432 cases that name a comparison model, 48.6% mention open or small models and 59.7% mention frontier models. -Criticism is common. Abandonment after trying Jev is not. 25.1% of all 6,277 cases contain criticism and 40.6% contain praise, but only 1 of the 2,153 hands-on cases ended with Jev being dropped. -Production use is still rare. Most activity is experimental. 80 of 2,153 hands-on cases (3.7%) are in production; prototypes, side projects and trial demos make up 62.4%. Seven days of data: treat this as a baseline, not a run-rate. Detailed Report fundaai.substack.com/p/deepl…

Preview| $MU FY26Q4: Focusing on Earnings Sustainability; A Higher, Longer Plateau Another beneficiary of emergence of Muse and other consumer agents Over the past three months, many market participants have been concerned that memory prices would drop rapidly after peaking, following the path seen in past consumer electronics cycles. However, we already expressed our view in our report following our attendance at FMS in early August: we believe memory prices are more likely to settle into a longer plateau after reaching their peak. We can broadly divide the memory market into two segments: consumer electronics and AI. In the consumer electronics segment, memory price increases are indeed approaching their limit; however, due to persistent supply shortages and limited newly added capacity, we see little chance of a sharp pullback in consumer electronics memory prices. In AI, while we clearly can no longer expect the QoQ increases seen in 26Q1 and 26Q2, based on our checks, we believe the price increases in 26Q3 and 26Q4 could exceed expectations and may continue to rise in 27Q1. Detailed Report fundaai.substack.com/p/previ…

Deep|LLM: Tiered Model Pricing Is Broadening AI Adoption; Limited Impact by Jev Pricing cuts accelerate adoption: Anthropic launched Claude Opus 5.5 at $4/$20 per million tokens, 20% below Opus 5, and estimates typical workload costs are down 40%. OpenAI’s GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50 are 50–58% below comparable GPT-5.6 prices, making more everyday AI tasks economical to automate. Tiered models expand use cases: Frontier models such as GPT-6 Astra and Claude Opus 5.5 still command premium pricing for complex tasks, while lower-cost models like Luna support extraction, routing, and consumer-agent workflows. The same application budget can now support more users, more model calls, or longer agent workflows. Jev is useful but narrow: TypeSafe’s Jev offers structured decisions at $0.042 per million input tokens with free output, averaging $0.0004 per call and 0.4-second latency in TypeSafe’s evaluation. It can replace some classification and guardrail calls, but it does not perform long-chain reasoning, so its impact on frontier-model demand should be limited. Compute demand remains supported: Over the next 6–12 months, broader AI deployment is expected to increase both usage and compute demand, even as inference efficiency improves. Application companies are the clearest beneficiaries, while model developers that compete mainly on low API prices face greater margin pressure, especially standalone Chinese players. Detailed Report fundaai.substack.com/publish…

The success of a personal agent product requires organizational commitment, user research, use-case iteration, application integrations, and access to CPUs and compute. $META has historically committed heavily to major consumer-product opportunities, and the required capabilities align with Meta’s strengths in consumer behavior and rapid product iteration.

This is exactly what we are doing. FUNDA's Context Layer Platform has already converted our own EPS Excel Models into database format, and everything can be handed over to our proprietary agent for precise modifications.

FUNDA has been conducting in-depth research on $INOD for some time. Last year, we published several detailed reports examining how META’s acquisition of Scale AI would affect INOD, drawing on extensive expert interviews and technical discussions. Deep|The Scale AI Halo Effect Updates: INOD May Haved Raised Its Sales Target This Year fundaai.substack.com/p/deept… Deep|The Scale AI Halo Effect: Will Meta's Investment Lift the Other Boats, or Just the Leader? fundaai.substack.com/p/deept…

In this note published yesterday afternoon, we summarise the checks that suggest that prices in 26Q3 and 26Q4 could exceed expectations and may continue to rise in 27Q1. $MU $SNDK fundaai.substack.com/p/previ…

We commented on this in our latest $INTC update: TSMC advanced nodes are fully booked, leaving Intel 18A as the most pragmatic base die partner for $SKHY & $MU.

Anthropic is facing its toughest competitive pressure right now, so it is not surprising that they are no longer in a rush to IPO in September or October. The best time would be after the release of Fable 5.2 / 5.5. At the current pace, Gemini 4, Fable 5.2, and GPT 6 SOL could all launch around the same time.

Someone asked what “recover unclaimed funds” means. When a company or institution owes someone money but can’t reach them and their account has been inactive for years, it generally has to turn the funds over to the state for safekeeping. Common examples include utility deposits never refunded after a move, balances in old bank accounts, uncashed paychecks, insurance payouts, and refunds from retailers. People can search by name and file a claim for free through official state websites or MissingMoney.com. The money generally remains available to claim indefinitely. Paid recovery services largely make their money by handling the paperwork and offering a service many people don’t realize is available for free. This is a good fit for Muse: the steps are straightforward, and most of the work can be done online. Muse already knows the user’s name and which states they’ve lived in, so it can search the relevant databases, check for matching records, and help fill out claim forms. Searching and filing are free, and users have very little to do themselves. That’s why people often report spending just five seconds or ten minutes submitting a claim. Those figures refer to the user’s time—not how long it takes for the money to arrive. Our tracking database contains roughly 20 such cases. About 15 include dollar amounts, totaling approximately $13,001 in identified funds. Individual amounts range from $12.85 to $4,700, with the three largest at $4,700, $2,900, and $1,750. The $1,750 was found by a user on behalf of his father-in-law.

Jev looks useful. We think the investment significance is being overstated. Plenty of software tasks need a label, a score or a yes/no answer. Jev handles those without generating text, at a very low price. There’s a market for that. The evidence supports a narrower claim than the excitement suggests. TypeSafe’s own benchmark measures agreement with larger models on four workflows. Independent results depend heavily on how the task is set up. Accuracy and calibration still need to hold up in production. For now, we see a niche product with attractive pricing and latency. We don’t see evidence that it changes the LLM roadmap or materially alters the economics of AI as a whole. Taking classification calls away from a small model is a long way from displacing frontier reasoning. Our investment view: no change to the AI infrastructure thesis. We would not revise training, inference or networking demand estimates on the back of this launch. We also wouldn’t underwrite a compute upside story just because cheaper decisions might create more usage. We don’t know the scale of either effect yet. For TypeSafe, cheap calls may get developers to try the product. Retention, production volume, margins and a durable quality advantage will determine whether there’s a valuable business. Worth trying if you’re building software. Too early to matter for the broader AI trade.

Most users are unlikely to complete the entire process. But even if they only go partway, Meta can capture purchase-intent signals and use them to serve targeted ads. That ad revenue has historically gone primarily to Google.