Neel KashCarry’s Biggest Fan 💯🖨💵

It’s becoming obvious to anyone with a chip design background that Nvidia is running into the physical limits of advanced packaging and quietly losing the technical moat that justified their valuation. Every new generation on their roadmap demands more design compromises than the market is pricing in. That means delays. And delays leave doors open for competitors who are already closing the gap on TCO. Exhibit A: Rubin Ultra already got gutted. The original four-die design had to be scaled back to dual-die because CoWoS-L couldn’t handle the thermal and structural stresses. The package was literally warping and compute dies weren’t making full contact with the substrate. TSMC’s advanced packaging hit a wall. That would be fine if Nvidia’s entire roadmap from Blackwell to Feynman didn’t hinge on pushing that same technology harder every twelve months. Their workaround now is a 2+2 configuration at the board level. But that’s a fundamentally different interconnect topology. You lose the die-to-die bandwidth density of a unified CoWoS interposer and fall back to PCB-level signaling between halves. For inference-latency-sensitive workloads, that’s a real performance degradation Jensen isn’t discussing and it’s exactly why he’s been so hostile to the idea of inference ASICs for years. He knows the cracks are showing. And my real fear is what if this isn’t the end of resorting to PCB packaging? Feynman is claiming 6 or 8 reticle dies on one interposer and they can’t even get 4 working? This would be absolutely catastrophic !!! Not to mention they even had to respin Rubin. After the first tapeout completed in late June 2025, Nvidia redesigned the chip in response to AMD’s MI450 and had to push TDP from 1800W to 2000W with a new tapeout not expected until late September/October now. That’s directly impacting the 2026 launch schedule and supply. But even more concerning than that is why does Nvidia need to make such big reactive design changes mid-tapeout on a chip this complex? It’s because the original spec wasn’t where it needed to be and their answer was to push more power to do it which isn’t a design win. On top of that, HBM4 verification delays at SK Hynix and Micron forced Nvidia to cut its 2026 Rubin production target from 2 million units to 1.5 million. This is the same movie as the Blackwell ramp: announce jaw-dropping specs, watch TSMC and memory vendors struggle to deliver, and let the revenue targets slip. The market doesn’t price Nvidia on delivering it prices it on announcements lol. Jensen knows this, which is why GTC has turned into a Hajj for speculators rather than a conference for design partners. So this is the core issue: the annual cadence of Blackwell (2024) → Rubin (2026) → Feynman (2028) is physically unsustainable but is necessary to justify its valuation. “Ultra” refreshes between each generation increasingly look like patches for unmet design objectives and every generation is stacking a new memory standard, a new packaging technology, and a new process node simultaneously. TSMC is already forced to explore alternatives like CoPoS (Chip-on-Panel-on-Substrate) because CoWoS-L can’t scale to where Nvidia needs it. All that introduces 6-12 months of delays and yield ramp risk and stacking all these is begging for even compounding delays. And this is the crucial point the market is starting to wake up to: when you have billions in data center construction costs on the line and customers to serve, you do not care where your chips come from as long as they meet spec and show up on time. It’s why Google has TPUs and Amazon has Tranium, the market for selling inference accelerators is much smaller than serving tokens… that’s the real business at the end of the day.

VP Vance left Friday for Pakistan: Axios

$app 1/.There’s a lot of emotion in the APP investor community, and there seems to be a real information gap between retail and institutional investors. Let me share what I’ve been hearing — hopefully this helps people get a clearer picture. This is not financial advice. 2/. Gaming is in healthy growth. There’s a “Q5” concept in gaming ad, and gaming clients are spending heavily into the “Q5” window. 3/. E-commerce is incremental: >50% of e-comm impressions are entirely new. The other portion is replacing low-price gaming bids, which is also net incremental. 4/. E-comm is off a low base but ramping quickly. The main category drivers this quarter are apparel, beauty, and subscription. Most are Shopify merchants — part of the reason is that APP did paid marketing in Q1. 5/. Discovery Campaign: after launched, many churned customers came back to test it. Customers ran a lot of incrementality tests and results were very good. Budgets are expected to continue ramping up in Q2. 6/. Prospecting vs. Discovery Campaign — both are net-new-customer acquisition products. Prospecting (launched 25Q4) can get new-customer ratio to 50%. Discovery (launched 26Q1) is even higher, >60%, which is why clients are ramping budget. 7/. Automated Creatives will help APP win more clients. Beyond video, APP’s ad formats also include end cards. If Meta needs ~10 creatives a day, APP needs roughly 15-25 or more (APP’s video ad is also much longer than Meta’s, which adds difficulty — Adam has said this many times). Agencies saw surging customer demand but can’t onboard them, and the reason is they don’t have enough creative headcount. As Automated Creatives keeps improving and rolling out, we should see more e-comm customers onboard. PS: an interesting stuff — I’ve seen some agencies already spin APP out into a standalone media team. 8/Meta: Meta’s “share gain” in late January only lasted 2 days (I think most people already know this). The reason was that Meta launched a more aggressive format and party AEM, but the results were underwhelming. 9/. Google/Unity: starting in 25Q3, I’ve heard Google/Unity has been offering gaming publishers incentive to migrate from MAX to AdMob (rough range few to ten millions, I can’t verify the exact number). So far I haven’t heard of any publishers migrating (could be source bias, but I think even if some have, it’s not material). Publishers feel that even though they’d get a chunk of cash short-term, it would hurt the company’s medium-to-long-term monetization, so they’re unwilling to leave MAX. 10/ On product iteration: customers say APP’s product team is willing to maintain high-frequency communication with them, which is how they’re able to keep shipping new products.

After considering all events in the region, and attitude towards negotiations past 48h, it appears that ceasefire is not taken seriously and that military operation is likely being prepared to take control of Hormuz (islands, and possibly Kharg), likely very soon.

*MICHAEL BURRY SAYS HE ADDED TO NVIDIA PUT CONTRACTS

Iran Parliament spox says no negotiations before 1) ceasefire in Lebanon 2) release of Iran’s blocked assets

Two of the measures mutually agreed upon between the parties have yet to be implemented: a ceasefire in Lebanon and the release of Iran’s blocked assets prior to the commencement of negotiations. These two matters must be fulfilled before negotiations begin.

VAHIDI HAS ALSO PUSHED FOR THE INCLUSION OF MOHAMMAD BAGHER ZOLGHADR, SECRETARY OF THE SUPREME NATIONAL SECURITY COUNCIL, IN THE NEGOTIATING TEAM, A MOVE OPPOSED BY CURRENT MEMBERS, WHO CONSIDER HIM LACKING EXPERIENCE FOR STRATEGIC NEGOTIATIONS