"Micron to buy back 30-50% of the company in the next 2-3 years" - UBS "Given FCF of ~$155B/~$225B/~$205B for C2027/2028/2029E, respectively, this implies the repo here could be in a similar ballpark as NVDA despite having an 80% smaller market cap. Said differently, this now lays the path for MU to buy back ~30% of the company over the next two years and potentially about half through C2029. Ultimately, we think the market will reward this stock with a much higher P/E multiple as it sees a cycle play out in which MU's gross margins probably don't go below 65-70%, but in the meantime TBV/share should approach $360 by CYE28 - keeping a solid floor under the stock given much higher RoE across the cycle going forward as memory is critically important in an AI world."

Broadcom CEO Hock Tan breaks down AI economics: Open-weight models burn $100B compute to make $30B in revenue. Frontier models spend $100B to make $120B in revenue. One of these won't be sustainable source: Goldman conference

Interesting call on the Google - $RDDT partnership with a former lead at Google, as well as why AI talent is leaving $GOOGL The read is that Reddit's negotiating position has deteriorated materially His framing is that the original agreement was a function of where large language models sat at that specific moment. Bard was becoming Gemini, hallucination rates were high, and the model had no grounding in current events. Training produces a model that is effectively six months stale by the time it ships. Reddit solved a narrow, acute problem: real-time human commentary at breadth and depth. The competitive set for that data was thin. X, Meta's properties and Threads are walled gardens aligned with rival frontier labs and were never gettable. Reddit was the one large corpus of unfiltered human language actually available for purchase, which is why both Google and OpenAI ended up there. The price reflects how little this mattered to Google's P&L. Roughly $60m: "For Google, $60 million to buy specific data is not a lot of money." His view is that the cash was the least interesting component. The valuable consideration was prompt data flowing back—what the user asked, and what they asked that led to a click through to Reddit. Both sides of the intent coin. He connects this to the trajectory of Reddit's advertising business, which has scaled well beyond what a new sales team and new infrastructure alone would explain. On the widely discussed point that AI Overviews traffic converts poorly, he broadly accepts it but argues the second-order effect dominates. Reddit held primacy in the citation slot, so volume was high even if quality was low, and the learning from that volume compounded. Google has since connected its search index and corpora more directly to the model layer, so the original grounding gap has largely closed. YouTube citation share has overtaken Reddit. Google News, Merchant Center and Places cover most of what Reddit was a shortcut to. His read on Reddit publicly signaling it might walk is that this is negotiation conducted through the press, and that it implies Google came back with worse terms—likely stripping preferences and the prompt data return rather than simply cutting the number. Google's standard posture on data rights is full and unrestricted use, and carve-outs on usage are not how Google contracts. "They probably don't need it. They probably want it." He expects a deal—the relationship is warm and mutually beneficial—but on terms that reset Reddit's expectations. Money is the secondary variable. The variable that matters to $RDDT holders is whether prompt-level signal keeps flowing. Asked whether Google would simply take the data if talks collapse, he says no on cultural grounds, that it is not in the corporate DNA to do that after the fact. More interesting is his description of how frontier labs behave when sued over training data: they do not settle, because a settlement establishes a price and invites every other rights holder. They litigate, spend, and drag it to a quiet resolution specifically to avoid setting precedent. Independent of the Google relationship, he identifies the harder issue: Reddit sits in the middle of a considered purchase journey with nothing to sell at the end of it. A user researches a bike on Reddit and buys it somewhere else. Two steps, and the second one is where the margin lives. The old funnel involved ten websites and fifty data points before purchase. That discovery phase is now collapsing into the chat interface, and the losers are the intermediaries that monetized the journey rather than the transaction. This is a Gemini problem, a Claude problem and an OpenAI problem simultaneously, not a Google-specific one. Search advertising worked because the system was deterministic—a tree you navigate from trunk to branch to leaf, ending in a transaction. Token predictors give wildly different answers to marginally different prompts, and the labs do not fully understand their own models' behavior post-training. Tuning that for advertiser ROAS is closer to dark arts than to keyword auction mechanics. The deeper constraint is grounding. Merchant Center is the largest product data repository in the world, Places the largest inventory of shops and locations, and every flight and hotel has been tuned within an inch of its life because advertisers were trained over two decades to feed that system. OpenAI has none of it. He is measured on the disruption question rather than dismissive. He cites Walmart attributing roughly 20% of traffic to OpenAI as evidence the relay is real, and he sees a credible alternative path: merchants exposing their own data through open interfaces that models come and fetch, rather than piping it into Merchant Center. That inverts the current architecture, but it needs an ROI case to bootstrap and there is a chicken-and-egg problem. His conclusion is share erosion and ad revenue siphoning, not collapse. On the suggestion that OpenAI should just acquire an ad tech engine, he is dismissive for the right reason: buying keyword-era infrastructure is buying an internal combustion engine in an electric vehicle world. A paragraph-long voice prompt carries vastly more intent than a three-word query, and the extraction method has to be built for that, not retrofitted. "Google has innovator's dilemma on steroids." A USD 250bn high-margin advertising business, powered by data that advertisers were trained to supply, cannot be hard-switched to Gemini. The chosen path is to infiltrate Search with AI and tolerate a messy middle until ROAS re-stabilizes on the new medium. He is candid that the current state is worse than Search at its peak, and he references the reported history of deliberately degrading search ad quality to increase monetization as evidence that the profit motive is explicit. On Google execution speed - Every objective must ladder from the most junior contributor up through director and VP. Search sits at the top of the tree, Android just behind, peripheral products have no influence. Strategy gets negotiated between silos, which is slow. "innovation is by definition outside of that data structure." Work outside the ladder is unsanctioned, and unsanctioned work costs you promotions and raises. He extends this to DeepMind, which he believes is now materially less isolated than it was and is being pulled into commercial delivery, and offers that as the explanation for researcher attrition to Anthropic and OpenAI. Not compensation—loss of research autonomy.

Good call on Bloom Energy $BE with a director on Crusoe's energy team TLDR: Constructive on $BE through the speed-to-power window: everything they can build gets sold, the regulatory-constraint use case is expanding, and the edge market is a genuine second leg. The bear case is not execution risk, it is terminal value—an unsubsidized, unconstrained power market where Bloom's cost curve has not moved enough. Key insights: Behind-the-meter adoption is not a technology preference, it is a hedge against binary regulatory risk. Moratoriums, temporary pauses and Governor Abbott's ERCOT announcements do not degrade a project's economics—they kill it outright, and approved power becomes worthless if a data center moratorium lands on top of it. He has seen two separate tenants elect to proceed with behind-the-meter projects alongside their grid portfolio, explicitly as diversification against that risk. Crusoe deliberately sites behind-the-meter gas away from rural communities to strip out both grid-connected regulatory exposure and community sentiment risk. Texas is the tell. It has historically been the easiest place in the US to get grid power because ERCOT is deregulated—no capacity market booking, no requirement to point specific generation at a specific load, with scarcity pricing left to incentivize the build-out. The fact that delays and regulatory uncertainty are showing up there is what has shifted tenant behavior. Working from a 100 GW / five-year data center demand frame, which he treats as aggressive but attainable: Six months ago: roughly 70 GW served by grid, assuming turbine production picks up, no fuel constraints, and several other unlocks. Some observers penciled 10-20 GW of SMR by end of period. Today: near-term grid share compresses to perhaps 50-60 GW, with the forward split moving closer to 50/50. Then it reverts. Once ratepayer protection is formalized, the grid reasserts as the cheapest and most reliable long-term supply. The mechanism for reversion is the Arizona construct APS has pushed—growth pays for growth, where all incremental upgrade costs are borne by the data center operator. He expects that to be cemented and formalized across every market, and expects it to take six to twelve months before politicians stop feeling their seats are threatened. The Bloom Bull Case He does not dispute near-term demand at all. "as much reliable Bloom capacity or solid oxide fuel cell capacity that can come online, will be deployed." If US deployable manufacturing produces 2-8 GW over the next two years, it gets absorbed. The premium is not a problem for buyers whose binding constraint is capacity. He also believes the product works. Hundreds of megawatts is deliverable, the systems are reliable, downtime is low because units are swappable, and the architecture is fully modular. The Bear Case Is About Price - "I don't see how they compete on price." Ten years out, in an unconstrained power market, he does not believe Bloom is competitive. He specifically does not believe the roughly 10% annual cost-down, on the grounds that the technology is structurally hard to make cheaper. He also flags a cost item he thinks the market underweights: the full system swap over a ten-year cycle, which makes ongoing O&M more expensive than a gas gen equivalent. Initial capex is the wrong lens; actual LCOE is critical to assess. He extends this into a coherent explanation of Bloom's own behavior. On the question of why they have not simply built the next facility if demand is as strong as claimed — his answer is not that they are sandbagging. He thinks they are building as fast as they can, and that the constraint is the supply chain, not the building. Solid oxide is an extremely small US market. You can put up a structure; scaling the full assembly chain behind it on the same timeline is the harder problem, unless more of it moves offshore. Crusoe Has Zero Bloom Projects Today - "Today, actually, we don't have any projects that are relying on Bloom fuel cells." The strategic rationale for staying at arm's length: "We've been a follower in this instance... we will accept it once the utility does." Buying 500 MW of Bloom directly means absorbing the regulatory risk that utilities may not accept solid oxide as high-rated ELCC capacity. Let the utility carry that. Turbines, recips, aeros and engines are all accepted technologies with known extreme-weather behavior and established effective load carrying capacity ratings. Solar, wind and battery now have them too. Solid oxide does not, because utilities have not yet observed large-scale fuel cell fleets through heat events and cold snaps. AEP has gotten comfortable off the back of the roughly 80 MW deployment plus smaller installations and Bloom's published test results. That is the template, and it is why the next few hundred megawatts of live operating hours matter far more than any order announcement. Where Bloom Actually Wins His model is not that Bloom wins on merit in a fair fight. It is that Bloom is the path of least resistance when a specific constraint blocks a project that already has hundreds of millions of development dollars sunk into it. If the binding constraint is price, Bloom loses. If it is speed to power, Bloom sometimes wins. If it is emissions, air permitting, noise or a regulatory restriction someone failed to plan for, Bloom wins. He reads the Nebius Vineland switch from gas gensets to solid oxide exactly this way — anti-genset pushback threatening a contract worth billions, with an obvious substitution available. He expects more of that, in lumpy project-specific chunks rather than as a smooth share gain. He also identifies the next leg of NIMBY-ism, which he thinks is underpriced: if communities dislike data centers, they dislike new gas generators considerably more. A fuel cell is lower emissions, quieter and a different class of asset. He calls it artful. That is a real, non-obvious tailwind. The Edge and Inference Market Is the Bigger Prize Crusoe is planning heavily for 10-50 MW modular builds, which he sizes at 20-40 GW over five years and would anchor at 20 GW in isolation. Amazon, NVIDIA, Tesla and xAI are all chasing the same edge market. "Bloom will leapfrog any gas combustion." In metro locations you cannot air-permit gas gen at all — this is a hard prohibition, not a noise preference. But he immediately caps the enthusiasm: much of that edge capacity is low-hanging fruit, converted Bitcoin sites at 5-18 MW with existing grid interconnects. "Grid power will always beat it." Bloom is the answer for incremental capacity and backup, not the base case. This is why he holds solid oxide at roughly 10 GW of the 100 GW mix, against 60-70% backstopped by combustion gas. He validates the native DC output argument, but for a better reason than efficiency. Fewer conversion losses lower the price, yes. The larger point is that it removes dependence on transformers and switchgear — equipment that is not only expensive but carries lead times that are themselves the binding constraint. A 345 kV breaker is two years out. Engineering around that bottleneck is precisely the kind of scenario where a project with sunk capital selects Bloom. What Changes His Mind He names two variables explicitly. Power price curves — if Bloom's costs do not fall, or if turbine pricing keeps rising, the relative position shifts and Bloom is in the money reasonably soon. And political sentiment, which he calls a big unknown and which he thinks is the more likely driver of fuel cell share than any technical milestone. Notably, a successful 200 MW deployment alone does not change his view. He already believes they can do it. source: Tegus

I halved my $GOOGL position: - They haven't released a new model in ages - Gemini Pro regularly crashes - A lot of the best people in AI just left - Search is 60% of profits, which is going to get largely disrupted - The parts I like are GCP and Youtube, but these are small parts of the business (35% or so) - 27x forward PE, fairly expensive given the risk in Search I've kept half for the moment on the hope they'll release a cool model soon, but all of this position is pure profit. I first invested in Google in 2011 and have gradually been selling over the past two years as Anthropic is just on a much faster innovation cadence and given my bearishness on Search. Anthropic is launching one impressive model after another. I do all my coding with Claude. Whenever I try one of these open-source models, they're garbage. It's called 'benchmaxing', as Palantir explained well - i.e. make the model look good on predefined benchmarks, but not being able to handle real workloads that engineers need. My impression is that Anthropic is running away with this and that the gap with competition will only increase from here. Anthropic has the cash to build custom data sets, has the talent, and the compute. Google seems to be lacking in execution. I don't want to write them off fully, but it certainly doesn't feel like high conviction.

Good expert call on Anthropic with a Head at Cognizant Cognizant ran GitHub Copilot, concluded it was not working, evaluated alternatives, and moved to Claude Code. That was announced late last year and is now rolled out broadly rather than sitting behind special approval. His view is that snapshot benchmarks are not the basis for a decision like this, because model leadership rotates. What drove it was that Anthropic is climbing the enterprise value chain faster than OpenAI, which he characterizes as having been model-first and then consumer-first. For a system integrator, the B2B orientation is the product. There is also a specific technical reason tied to their core delivery work. Cognizant does heavy legacy modernization, pulling apart COBOL mainframe estates. Claude Code's larger context window lets them work on bigger chunks of a migration at once, which is the binding constraint in that work. "we are just half a step ahead of the customers and maybe not even that" Developers are running 30-35% more effective. That is spreading into testing, and then into business analyst work, requirements capture, prototyping and UI drafting, with architecture behind that. The commercial problem is that an integrator on time and materials does not automatically capture its own efficiency gains. He places this inside a 10-15 year erosion of pricing power: clients have become far more professional buyers, and the platform companies are taking a growing share of total IT spend. Outside of engineering he describes it as very early. Most people are drafting emails and getting help with presentations. The step change he describes is a project brain: an entire RFP, with notes, templates and background, loaded into a single Claude project that a whole team works against. There are also hard limits on what can be uploaded, and customer contracts cannot go in. He is specific about what agentic automation needs before it clears the bar for regulated back-office work: governance, happy path analysis, human-in-the-loop, auditability for BPO processes, confidence scores, and an ontology over the underlying documents so that a policy document is parsed the same way every time. His view is that Anthropic does not have that functionality yet. For live BPO work they use a native agentic automation platform instead, Otera, which is Austrian and fully EU-resident including data centers and personnel. He expects Anthropic to get there on the trajectory of the last six months. The gating factor is not the software. "the end state will make the current consumption look like nothing... it's a business transformation, and it needs to be driven like one" He explicitly guards against recency bias. Everything has gone Anthropic's way for five or six months and he thinks the market is over-extrapolating that. His view is that Anthropic is pulling ahead but it is early, that OpenAI will have a real enterprise play, and that the space comfortably supports two primary providers. He is more concerned about the second tier, naming Meta, falling behind. source: Tegus

"The limiting factor currently is memory. The memory output is increasing by around 20% per year. Now normally, that would be fantastically fast and amazing for any large mature industry. But, the demand is increasing by 200% a year, maybe higher. So if you've got demand increasing much faster than supply, then Economics 101 would suggest that the price increases. It does not decrease." - @elonmusk Yes, I'm still long DRAM $MU $SKHY

High quality call. The expert spent 31 years at $INTC, 24 of them inside the fabs, running 24/7 operations across lithography, dry etch and thin films metals as an area manager. He traces Intel's derailment to one call: for 10nm, Intel decided it did not need EUV, despite having been one of the largest investors in ASML to create the technology. The company committed to technical specifications that could not be delivered reliably without it, then tried to hit them with quad patterning. "Despite Intel being one of the biggest investors in ASML to create EUV technology a decade earlier, by the time that technology was ready, we said, 'Nah, we don't need it.' That was a watershed decision... Yields, costs, throughput, all of it suffered." The mechanism that turned a process error into a company-level failure is the IDM coupling. Intel products depend on Intel Foundry. When 10nm underdelivered, the product line had no competitive process to design onto, and the only remedy was outsourcing, which is why a meaningful share of Intel's chiplets are built at TSMC. Low internal volume then starved the fabs of the very thing they needed to recover. His framing of the node sequence is bracing: Intel 7, Intel 4, Intel 3 and 18A were a chain in which each node inherited the unresolved problems of the one before, and 18A arrived initially without yield or parametrics. He puts the effective slip on 18A at roughly two years. Process improvement is a data-limited exercise. TSMC runs full loading for Apple, AMD and NVIDIA, which generates enormous volumes of process data, which enables fast learning cycles, which sustains leadership, which attracts the volume. Intel is fabbing almost exclusively for Intel products, and Intel products have been weak for a decade. The data turns are not there. "They're basically only fabbing for Intel products, and Intel products have been struggling. The volume of data is just not there for Intel to innovate faster... I think sometimes the reason Intel hesitates, the reason they waffle a little bit on making claims is because of this, because they just can't get the data turns fast enough to know with confidence where they're at." He flags Foveros and wafer-level packaging as equally critical to 18A and increasingly the harder problem. Panther Lake does not ship without it, and a packaging failure destroys yield at a far higher cost per unit than a front-end failure, because you are scrapping fully processed die. Lip-Bu Tan's original position was that Intel would not commit to 14A without external customers putting skin in the game. The expert's point is that this position was internally incoherent with the stated commitment to remain an IDM. If Intel does not build 14A, Intel products cannot design competitive parts, so Intel products must be allowed to outsource wafers. Do that for a couple of years and the foundry loses its largest customer and is finished. He also thinks the original statement was a strategic error in communication terms, since no prospective customer is motivated by a vendor signaling it might stop investing. "Intel products cannot design competitive products if the foundry is stuck on an old node. If Lip-Bu stuck to that initial declaration, he is then required to allow Intel products to outsource their wafers. By default, over time, not very long either, within a couple of years, he puts his own foundry out of business, because his number one customer is Intel products." He knows that prospective customers have been running test chips, samples and data reviews with Intel Foundry for at least two years. The absence of commitment after that much engagement is itself the datapoint on process competitiveness. What could change it, in his view, is not purely technical. Two exogenous tailwinds are now in play. TSMC leading-edge capacity is on allocation, so second-source conversations are being driven partly by desperation rather than preference. And there is a political layer around domestic capacity that he assumes is present in the shadows of these decisions. He credits Lip-Bu Tan with identifying waste and making necessary decisions into a genuinely broken situation. His objection is to method rather than direction: capable long-tenured people pushed out because the incoming CEO did not know them, replaced by people brought in from adjacent industries. He also puts the start of the decline around 2012, well before Gelsinger, and characterizes Gelsinger's plan as broadly right but poorly executed and over-committed. Conclusion: Intel is obviously far behind TSMC when it comes to process know-how and yields, but has an opportunity in the coming years to pick up volumes as TSMC is at 100% utilization. Larger volumes should lead to better yields and increase the learning curve for Intel, but obviously, this remains a difficult story. TSMC is obviously far ahead and is building additional N3 and N2 capacity to address wafer supply shortages themselves. $INTC $TSM $ASML

Deutsche expert call on humanoid robotics. Positive indicators already: >2 years payback, 80-90% success rate. Clearly, this is not something which is going to have an impact on any of the semi names in the coming years, but long term the potential is massive here.

Nice expert call on $CBRS with a Director at Figma who's sending 40% of AI traffic to Cerebras' cloud There are clear use cases for hyperfast inference and the customer is willing to pay a premium for these fast tokens Figma spent roughly a year building an in-house language model for its assistant product, a chat interface intended to make Figma AI-native. The model is in the 50 to 100 billion parameter range, which he characterizes as roughly five to six times smaller than the frontier models it replaces, and better in quality for the specific task. Figma ships the weights to Cerebras and Cerebras handles all serving and optimization. He explicitly frames the relationship as a cloud provider relationship, not a hardware purchase. The AI traffic split: - 20% of assistant traffic routes to external frontier models (Anthropic and others) for general question answering and RAG-type requests. This share is structurally permanent. Figma has no intention of training its own model for it. - 80% routes to the in-house model. - Of that 80%, half goes to Cerebras, so roughly 40% of total assistant traffic. - The remaining 40% goes to Fireworks, Modal and Baseten. Figma Make, the vibe-coding product, sits in the same org but runs almost entirely on external models from Anthropic and Google. Critically, the AI assistant is currently exposed to only around 30% of the Figma user base. Current Cerebras spend is described as low to mid seven figures annually. His forward view: "I would expect at least in the next one year to easily double... maybe 2X-3X growth in the next one to two years." Beyond that he expects the use case to plateau, since the core objective is penetrating the existing user base rather than acquiring new users. This is a useful anchor for how a mid-size, non-AI-native enterprise customer scales on this platform: fast doubling off a small base, then aiming to flatten once penetration completes, with the next leg dependent on new products rather than the same product growing. # Why Cerebras Does Not Get the Other 40% of Traffic 1. Capacity is reserved, not metered. Figma has weekday peaks and effectively zero weekend and overnight traffic. Reserving for peak means paying for idle silicon most of the week, so Figma deliberately sizes its Cerebras reservation at only 40% to 50% of average traffic and sends the peaks to on-demand GPU providers. This single design choice is what creates the opening for Fireworks, Modal and Baseten. 2. Reliability - Error rates are still above what he wants from a production system, which forces Figma to maintain fallback paths. 3. Observability - He wants first-try success rates, retry counts and internal failure data, which he is not currently getting, and without which he cannot engineer around the reliability gap. "If Cerebras did not have any limitation, we would've used the whole 80% of traffic through Cerebras." # The Deployment Friction, Scored 7 out of 10 Model updates are not self-serve. Figma has to notify the Cerebras team roughly a week ahead. Hand a model over Monday or Tuesday, and it is deployed by Friday. Two to three days per iteration, versus effectively immediate self-serve deployment at every GPU-based competitor. He rated the pain at seven out of ten on a scale where ten is severely damaging. "It's of course something we can live with, but it definitely slows our execution a lot... we build a model, we do some testing, we do this three-day wait, do some testing, find that there is a small bug, and then we have to retrain the model." This is a developer infrastructure maturity gap, and it is the kind of thing that does not show up in benchmark comparisons but does show up in renewal conversations and in how much of a customer's roadmap a vendor can capture. # The Price of Speed Against the frontier model, running Figma's own smaller model on Cerebras costs roughly half to one third as much for equivalent traffic. But the model is five to six times smaller, so the like-for-like inference premium is real. Against a GPU-based provider running the same model, his estimate is that Cerebras is around 30-50% more expensive. Figma has not run a formal side-by-side, which he attributes to the fact that reserved pricing versus per-token pricing makes the answer entirely dependent on the traffic profile. In a bake-off against hyperscalers, Modal, Baseten, Fireworks and Groq, Cerebras came out roughly 10 to 15 times faster than most of the market. Figma found Cerebras through the Artificial Analysis public benchmark, then spent months on POCs specifically to de-risk infrastructure maturity before scaling to production. The latency threshold he describes is a genuine product constraint rather than a nice-to-have. For assistant tasks a user could perform manually in about 30 seconds, an AI response taking longer than that adds no value. Delivering in three to five seconds changes the product. "Speed is extremely valuable... even that 50% to 2X increase in price, I think is totally something we are willing to pay for the speed." Cost optimization at scale would likely not come from switching hardware. It would come from shrinking the model further or routing simple requests to a smaller model. "Maybe a workflow user makes a request, it takes three hours to run. If that's the case, then speed is not a big criteria... For those use cases, we may switch to the cheaper inference provider." Figma has not built an in-house model for Figma Make because the output format is React and HTML, which frontier models have seen extensively, leaving little quality headroom. But he notes cost and latency wins are still available, and if cost becomes a concern as Make scales, Figma would likely train a smaller model and host it on Cerebras for the base load. Additional projects beyond the assistant are early stage, with some expected to scale in the second half of the year. He also notes Figma experimented with Groq prior to its acquisition by NVIDIA. $CBRS $NVDA