Why is NVIDIA acquiring Hugging Face? It comes down to a simple match: one side has the best hardware on the planet, and the other has the community. Hugging Face is where 18 million developers build open AI, and having massive tech power behind them means everything will run way faster and smoother. It makes free AI models easier for everyone to use and gets new tech into everyday apps much quicker.

$WYFI is easily one of the most interesting neoclouds on the market right now. Instead of overleveraging their balance sheet just to buy GPUs, they are building a unique, capital-light model - combining durable colocation assets with partner-led sites, and high-margin managed infrastructure services.  Massive credit to the @MarkosAAIG and @Mark_AAIG for such a thorough, rigorous, and deep strategic breakdown. It does a fantastic job of dissecting the underlying logic of the business.  If you are a shareholder or considering buying $WYFI, this is a total must-read.👇

$HPE earnings tonight after the bell. Most investors look at $HPE and see AI servers. After yesterday’s $DELL results, it’s easy to focus on backlog growth and AI infrastructure demand. But I think the real story is Networking. With Juniper now part of HPE, the company has a much stronger position in one of the most important layers of the AI stack. And judging by the current valuation, I don’t think the market is giving that business nearly enough credit. Here is what I’m watching across their 3 main pillars and why this trade plays out on multiple fronts: 1. Cloud & AI Expecting nothing short of phenomenal numbers here. Following $DELL great earnings, we already know how insane AI demand is right now. I’m expecting a solid beat, further expansion of their already record AI backlog (previously $6.3B), and a guidance hike for the full year. 2. Networking + Juniper Networks HPE isn't just selling switches - they now directly challenge Ciena and Cisco (+25% op margin) in high-margin optical networking, coherent optics, and Data Center Interconnect (DCI), while targeting Arista Networks (43% op margin) in AI fabric. HPE’s standalone networking sits in the low 20% range. I ntegrating Juniper’s optical/routing tech and Mist AI provides a complete stack (from AI servers to optical interconnects), driving structural margin expansion. The market trades pure-play networking at a massive premium. As integration progresses, HPE gets re-rated from a low-margin hardware vendor to an enterprise networking power. 3. Hybrid Cloud & GreenLake (The Cash Flow Anchor) Traditional IT and the GreenLake subscription model provide steady cash flow to buffer hardware margin swings. This segment generates the dependable capital needed to rapidly pay down debt and deliver on their $3.5B+ Free Cash Flow target for the year. Wall Street is modeling Q3 revenue around $11.5B–$12.1B and EPS at $0.88–$0.93. With expectations set high after Dell, an AI beat is table stakes - the real upside trigger will be networking integration velocity and margin resilience.

During yesterday's earnings call, $DELL laid out exactly what's hitting the supply chain red lines. We have a shortage of: > Memory: DRAM, DRAM, DRAM, followed by NAND, NAND, NAND > CPUs: Spotty CPU shortages > Storage: Disk drives > Foundry capacity: Leading node products > Mature nodes: MOSFETs, power ICs, microcontrollers, and drivers > Packaging & materials: ABF Substrates and Tea Glass > Networking: Optical components > AI infrastructure: CDUs (Cooling Distribution Units) and power racks Hey, is there anything left out there that ISN'T in short supply? What a times. Any suprised?

Quick thoughts on $CRDO after their earnings call 👇 ➡️Optics is taking the crown from copper AECs built the Credo we know today, but optics is officially taking over as the primary growth engine. Credo reiterated >$600M in optical revenue for FY27, with ZeroFlap optics, PICs, and optical DSPs each contributing >$100M. They are successfully moving past the "just a cable company" narrative. ➡️Dust Photonics looking like a brilliant M&A play The Dust Photonics acquisition paid off immediately with its first recognized PIC revenue in Q1 and confirmed FY28 design wins at 2 major players. Pairing in-house silicon photonics (PICs) with their custom DSPs gives Credo a clear cost (COGS) and system-level performance edge over standard module builders. ➡️Fixing one of AI’s biggest bottlenecks (ZeroFlap + Pilot) Link flaps in massive AI clusters can waste 10% to nearly 20% of expensive GPU utilization. Credo’s Pilot software acts like a "check engine" light—spotting fiber dust and latent ESD damage before a link actually fails. Slashing cluster bring-up time from 6–8 weeks down to just 5–6 days is a massive value prop for hyperscalers. ➡️Concentrated hyperscaler traction (and product expansion) Customer concentration remains high - their top 4 customers drove 84% of Q1 revenue (33%, 28%, 13%, and 10%). However, Credo now holds deep relationships with 5 top hyperscalers overall and is expanding across Neo Clouds. Crucially, these Tier-1s aren't just buying AECs anymore; they're adopting Credo's broader optical and retimer portfolio. To support this wave, Credo has been prepping its supply chain for 18–24 months - leaning heavily into working capital with inventory up $62M+ sequentially to $313M to ensure they can scale production effortlessly for the 2H FY27 ramp. ➡️Sneaky monetization play on Inference (OmniConnect) Credo is positioning early for the memory fanout and bandwidth wall in inference. Their Weaver gearboxes interface with LPDDR5/LPDDR6 to boost memory capacity up to 2TB (starting with customer Positron). Management expects this to represent thousands of dollars in Credo content per GPU starting in FY28. ➡️Agnostic strategy over religious tech battles Instead of picking sides in copper vs. optics vs. NPO/CPO, Credo is playing the whole board: > Short-reach copper (AECs up to 6.5m for 1.6T) > Micro-LED active cables (ALCs up to 30m, ramping FY28) > Full optics & NPO (joined Open CPM MSA consortium for scale-up) They win regardless of which topology hyperscalers or Neo Clouds choose. What’s your take on $CRDO here?

The $1T number is the part I’d focus on. AI infrastructure is becoming a multi-trillion-dollar capex cycle. GPUs get the headlines, but the infrastructure around compute like power, networking, cooling, memory and storage - could be some of the biggest beneficiaries. I think we can agree this isn’t a bubble.

$DELL latest numbers look massive, but the actual story is all about profit margins. They raised their full-year revenue target to $192 billion. Sounds great, right? The catch is that lower-margin AI servers now make up over a third of their sales, which naturally squeezes their bottom line. They pulled in $60.9 billion in AI orders, pushing their total backlog to $95 billion. The demand is clearly real, but turning those orders into delivered products without hurting profitability is going to be tough. The actual win for them this quarter? Traditional servers grew 122%, and storage went up 26%. That higher-margin gear pushed their ISG operating margin up to 15%. Is Dell building a long term business model here, or will low-margin hardware pull them down later?

Just as I predicted - a classic beat & raise for $CRDO Credo blew past Wall Street estimates across the board: revenue hit $479M vs the $435M expected, and Non-GAAP EPS crushed it at $1.20 vs $0.98 consensus. The Q2 guidance of $525M–$535M completely wiped out analyst expectations ($465M). AI infrastructure demand is real, Non-GAAP margins hold strong at 68%, and operational leverage is insane. Wall Street will be forced to aggressively hike price targets after this clear execution showcase. How market react? Of course - sell out, what the times.

Today after the close, $CRDO reports earnings. I don’t have a crystal ball and I have no idea how the market will react. But I can see what’s happening across the industry, where the money is flowing, and where demand remains incredibly strong. In my opinion, Credo will beat expectations today and raise guidance for both this year and next year. We’ll find out soon enough.

Samsung's zHBM was one of the more interesting things I came across today. Today, HBM sits beside the compute die. zHBM flips that idea and places memory directly on top using Hybrid Copper Bonding. The potential benefits are obvious: > Shorter interconnects > Higher bandwidth efficiency > More memory in the same footprint Samsung is talking about up to 8x performance vs HBM5, 3x better energy efficiency, and 10x higher memory density. Still a concept rather than something we're likely to see deployed at scale anytime soon. My guess would be 2029–2030 at the earliest. For now, it's an interesting look at where HBM could go next.

Good to hear a rational view on the impact of AI data centers on local communities instead of the constant FUD. But to be fair, let's look at both sides. The man speaking in this Fox News segment is a Loudoun County resident. He probably has a better understanding of what DC actually mean for a local community than most people who have only seen the letters D and C next to each other on a keyboard. And Loudoun is not exactly a random example. Ashburn, in Loudoun County, Virginia, is home to "Data Center Alley" - one of the largest concentrations of data centers in the world. Data centers have become an enormous source of local tax revenue and a major part of Loudoun's fiscal base. The economic case is pretty straightforward. Data centers generate hundreds of millions of dollars in local tax revenue They have helped support lower residential property tax rates They fund schools, public services and infrastructure Construction creates substantial demand for skilled labor They strengthen America's digital and AI infrastructure For FY2027, Loudoun expects roughly $1.3B in tax revenue from data centers, representing around 45% of projected county tax revenue. That's not a marketing talking point. The county itself documents the extraordinary fiscal contribution of the data center industry. And Loudoun isn't the only example. Fox News recently reported on Quincy, Washington, where data-center tax revenue helped fund a new high school, hospital and aquatic center. At the same time, Fox makes an important point: a single data center may employ relatively few people once construction is complete. So what are the legitimate arguments against them? 1. Noise Yes, this is a real issue. But it is primarily a question of engineering, site selection and regulation. AI facilities are increasingly moving toward direct-to-chip liquid cooling because of the power density of modern AI racks. This reduces the cooling-related airflow burden, although it doesn't make a data center silent. And compared with many industrial facilities operating 24/7, a well-designed data center can be a relatively quiet industrial neighbor. The relevant question isn't simply: "Are data centers noisy?" It's: "What is the measured noise level at the property boundary?" That's something that can actually be regulated. 2. Transmission lines Another argument that needs some nuance. A data center doesn't simply "consume" a transmission line. It creates a very large electricity load that requires sufficient generation, transmission and distribution capacity. Increasingly, power availability is one of the key factors determining where new AI data centers can actually be built. Developers generally pay their direct interconnection costs and can be required to fund or contribute to grid upgrades needed to serve their load. So the question isn't simply: "Do data centers pay for the grid?" It's which costs are directly assigned to the project and which are recovered across the broader rate base. 3. Electricity consumption This one is absolutely real. AI data centers consume enormous amounts of electricity and demand will continue to rise as compute density increases. But the question isn't whether America will "run out of electricity." The real question is whether generation, transmission and distribution can expand fast enough to keep up with new load. If they do, large new loads can help justify investment in new generation and grid infrastructure. If they don't, electricity prices and grid constraints can become an issue. Recent research found evidence of both effects: data-center entry was associated with higher local employment, wages and business formation, but also higher electricity prices in some markets. The distinction matters. It's not simply about how much electricity AI consumes, but how the system responds to that demand and who pays for it. 4. Water This is where the technology is changing the discussion quite quickly. New high-density AI facilities are increasingly adopting direct-to-chip, closed-loop liquid cooling, which can eliminate evaporative water use for cooling. And this isn't just a Microsoft story. Oracle says the AI data centers it is building today in several U.S. states use closed-loop, non-evaporative cooling. Google introduced its Brazos liquid-cooling system in 2026, while NVIDIA's latest Rubin architecture is designed around 100% liquid cooling. So: "AI data centers require huge amounts of water" is becoming an increasingly incomplete statement. Water use hasn't disappeared across the industry, but the cooling architecture of new AI facilities is changing rapidly. 5. Land This is a legitimate issue. If an AI campus means clearing valuable farmland or natural land and building an entirely new power and infrastructure network, there is a reasonable question about whether that is the best location. But a brownfield site with existing power, fiber and industrial infrastructure is a very different proposition. The question shouldn't simply be: "Does the data center use a lot of land?" It obviously does. The better question is: "Why was this particular site chosen?" 6. "They create very few permanent jobs." Also true. An AI data center isn't a car factory. Once construction is complete, relatively few people may work inside the facility. But measuring the economic impact only by people employed inside the building misses a lot. Someone has to: > build it >maintain it > service the electrical infrastructure > provide cooling systems > provide security > upgrade the equipment > supply networking hardware > provide engineering and construction services And there can be broader local effects. Recent research on data-center entry found measurable effects on employment, wages and business formation, with stronger effects in metropolitan areas where agglomeration effects are more pronounced. The same research also found evidence of higher electricity prices in some markets. That doesn't mean every individual data center creates thousands of local jobs. It means the economic impact shouldn't be measured solely by the number of people sitting inside the building. So what's the actual answer? I don't buy either extreme. "AI data centers have no impact on communities." Wrong. They consume huge amounts of electricity, require infrastructure, occupy land and can create real local externalities. But: "They are giant warehouses that consume power, water and land while creating almost no economic value." That's also wrong. The economic benefits in places like Loudoun are measurable. The right question is much more boring and much more useful: Is this particular project well located, and is the developer paying an appropriate share of the infrastructure costs it creates? If yes, an AI data center can be an extremely valuable economic asset. If not, residents have every right to push back. That's a much more productive debate than simply being pro- or anti-data center.

$MRVL FY2028 revenue target keeps moving higher: Sep 2025: $13B Dec 2025: $15B May 2026: $16.5B Aug 2026: $18B That's a 38% increase in less than a year. And we're only halfway through FY2027. FY2027 revenue guidance is now $12B. With Q3 guided at $3.15B, Marvell needs just $3.69B in Q4 to hit the $12B target. Then comes FY2028: $18B. That means the next 6 quarters represent $24.84B of revenue. Using Marvell's current 31.6% non-GAAP net margin, that's $7.85B of non-GAAP earnings. At $216.62, MRVL is trading at roughly 24.2x earnings generated over the next 6 quarters. The interesting question isn't whether management misjudged demand. It's whether AI infrastructure demand is evolving faster than long-term models can keep up with. $18B is today's FY2028 target. Will it still be $18B after the next update?

This is an interesting $AAOI update from @JonahLupton, but one thing in particular caught my attention: “Demand is not the problem. Capacity is.” That’s a very important distinction. If hyperscalers are already willing to sign LTAs and AAOI is saying it could win additional business if it simply had more capacity, then the $600M ATM starts to look less like “raising money because the business needs it” and more like raising capital to remove the bottleneck. And I think there’s another piece here that the market may be underestimating. AI infrastructure is becoming increasingly optical. Higher GPU density, faster networking and the transition to 800G/1.6T all mean more optical content moving through each generation of AI infrastructure. So AAOI doesn’t necessarily need to win some massive share of the AI market. It needs to keep increasing capacity while the amount of optical infrastructure being deployed keeps growing. That’s the part of the thesis I find most interesting. The numbers are obviously where this gets crazy: $1.1B revenue in 2026 Then $4B in 2027* *Potentially $4.5–5B if capacity ramps faster And potentially $8–10B in 2028 if the growth trajectory continues. This HUGE! But of course execution is the key risk, but if AAOI can actually build the capacity to meet this demand, I think the upside is enormous. And honestly, after connecting the dots, I completely agree with @JonahLupton’s $500–600 target over the next 18–24 months is very realistic. The market may be underestimating what happens when the bottleneck is not demand, but capacity. This isn’t just a $AAOI story either. $LITE and $COHR are saying essentially the same thing - everything they can produce is already sold out. And I think $AAOI will be the one to ramp capacity the fastest.

This is an interesting $AAOI update from @JonahLupton, but one thing in particular caught my attention: “Demand is not the problem. Capacity is.” That’s a very important distinction. If hyperscalers are already willing to sign LTAs and AAOI is saying it could win additional business if it simply had more capacity, then the $600M ATM starts to look less like “raising money because the business needs it” and more like raising capital to remove the bottleneck. And I think there’s another piece here that the market may be underestimating. AI infrastructure is becoming increasingly optical. Higher GPU density, faster networking and the transition to 800G/1.6T all mean more optical content moving through each generation of AI infrastructure. So AAOI doesn’t necessarily need to win some massive share of the AI market. It needs to keep increasing capacity while the amount of optical infrastructure being deployed keeps growing. That’s the part of the thesis I find most interesting. The numbers are obviously where this gets crazy: $1.1B revenue in 2026 Then $4B in 2027* *Potentially $4.5–5B if capacity ramps faster And potentially $8–10B in 2028 if the growth trajectory continues. This HUGE! But of course execution is the key risk, but if AAOI can actually build the capacity to meet this demand, I think the upside is enormous. And honestly, after connecting the dots, I completely agree with @JonahLupton’s $500–600 target over the next 18–24 months is very realistic. The market may be underestimating what happens when the bottleneck is not demand, but capacity. This isn’t just a $AAOI story either. $LITE and $COHR are saying essentially the same thing - everything they can produce is already sold out. And I think $AAOI will be the one to ramp capacity the fastest.

Great article from Jim breaking down the $IREN earnings. One detail I think deserves more attention: IREN explicitly highlighted twice that its current data center portfolio is unencumbered. I think this is more important than it may initially sound. Over the last year, IREN has used equity capital, customer prepayments and GPU financing to build out productive infrastructure and turn secured power into revenue-generating AI assets. Now those assets can potentially become collateral for a new layer of financing. That changes the capital structure quite significantly. Instead of continually funding growth primarily through equity, IREN can increasingly finance the next stage of its buildout against an existing asset base and contracted cash flows. And this may also explain the preference for medium- and long-term contracts: predictable contracted cash flows are much more financeable than short-term or merchant revenue. So the sequence matters: equity / customer capital → productive assets → contracted cash flows → asset-backed financing → more infrastructure → more contracted cash flows. That is the capital flywheel Dan Roberts was talking about on the earnings call. To me, this is one of the more important pieces of the earnings that the market may be overlooking.

$IREN plans to reach $4B ARR by the end of 2026. Now imagine 2027. If they add just 500MW of new capacity and each MW generates $20M per year, that’s another $10B in ARR. At $25M per MW, it’s $12.5B. Add the existing $4B, and we’re looking at roughly $14–16.5B ARR by the end of 2027. So when I wrote about $15B ARR for $IREN by EOY 2027, I wasn’t making it up. And here’s the crazy part: $IREN has a market cap of only ~$15B today. Think about that for a second.

2028 is still a long way off, but $IREN is already preparing for 800V DC today. That could become a huge moat. I think 2028 could be the year when $IREN really starts to separate itself from the competition.

MLCCs (Multilayer Ceramic Capacitors) - tiny, low-cost passive components are emerging as a critical, underappreciated bottleneck in scaling AI data centers. The move to new architectures (GB300 and beyond) sharply increases power density and component requirements: > MLCC cost on Compute PCB: $25 → $90 (+260%) > On Switch PCB: $20 → $45 > Total MLCC content per server rack: $1,530 → $4,320 (+182%) In high-end racks this means hundreds of thousands of individual capacitors. Demand outlook (structural supercycle): AI data centers are expected to drive >4x growth in high-capacitance MLCCs (>47µF) from ~4 billion units in 2025 to 38 billion in 2030. Supply grows only 10-15% annually → lead times >20 weeks and shortages in premium segments. Prices for advanced server-grade MLCCs are already up double-digits YoY. This shifts MLCCs from cheap commodity to strategic resource in the entire AI stack . Best positioned players (advanced miniaturization + high-capacitance + server exposure): Murata Manufacturing ( $MRAAY / -> holds a 40-45% global MLCC market share, which scales up to 50-60% in premium server and AI applications. It is the undisputed technological benchmark in material science, excelling at extreme miniaturization while maintaining high capacitance, giving it unmatched pricing power in high-end enterprise infrastructure. Samsung Electro-Mechanics ($009150.KS) -> holds a 20-24% global market share, capturing an estimated 35-40% of the AI server deployment volume. Backed by deep semiconductor packaging expertise, it is rapidly closing the gap with Murata in hyperscaler data center deployments. Taiyo Yuden ($TYOXY / -> holds a 13-15% global market share. It is a pure-play high-capacitance specialist heavily oriented toward power-smoothing and decoupling needs in massive compute clusters and enterprise infrastructure. Yageo Corporation ( holds a 12-15% global market share. A massive-scale multinational, it successfully penetrated premium tier-1 enterprise and server markets, stepping in where primary vendors face supply constraints. TDK Corporation ($TTDKY / holds a 10-12% global market share. Focusing on extreme-reliability, high-voltage, and high-temperature performance, its components are indispensable around the complex power delivery networks of high-wattage accelerators. 👇 Which of these MLCC leaders do you see as best positioned for the AI supercycle? Or which part of the passive components supply chain do you think is most underestimated?