Guy Resheff

Nvidia’s plan to restart H200 shipments to China by mid February reads like a small operational update, but it quietly reflects how complex the semiconductor landscape has become. On paper, this is about compliance and product adjustments. In practice, it shows how difficult it is to separate technology, markets, and policy once an industry reaches this level of global interdependence. From years of working across technology and capital cycles, this kind of move feels familiar. Export rules may change the shape of a product, but they rarely erase demand. Companies adapt, redesign, and recalibrate, not to challenge policy, but to keep long term relationships and ecosystems intact. That tension between staying within the lines and staying relevant is now part of everyday decision making. What’s interesting is not just Nvidia’s execution, but what it says about the broader system. Policymakers are trying to set boundaries, companies are trying to preserve continuity, and markets are trying to price in both. None of those forces are wrong, but they are increasingly out of sync. Over time, the gap between intent and outcome becomes harder to ignore. For investors, this is a reminder that control in semiconductors is rarely absolute. Capabilities diffuse, knowledge travels, and innovation finds paths that regulation did not anticipate. Adjusted products may look like compromises, but they often become bridges that keep the system connected. It leaves an open question worth thinking about. Are these moves helping stabilize a fragile balance, or are they simply buying time while the industry quietly evolves around the rules. #Semiconductors #Nvidia #AIInfrastructure #GlobalSupplyChain #TechnologyPolicy #StrategicInvesting #InnovationDebate

1

"The crunch is already inflating the cost of AI infrastructure and everything else that relies on memory. It threatens to squeeze margins, delay product launches and potentially render some devices too unprofitable to make...Data center demand for DRAM surged to around 50% of global consumption in 2025, up sharply from 32% 5 years earlier. That share is expected to climb further. By 2030, AI servers are projected to account for >60% of global consumption." As I travel around meeting with clients and prospects, few questions come up more frequently than the long-term viability of the genAI business model. It's something I dissected in depth in my report on "GenAI & Productivity" (https://lnkd.in/gEYKVHpU). It's something I'm particularly concerned about today given the Iran war triggering risk-off sentiment and stagflation fear. I often go back to a quote from Verdad's Dan Rasmussen: "What’s interesting about AI is that it’s the first tech innovation that’s capital intensive since fiber…Big tech companies have gone from having about a third of the capital intensity of US industrial companies to three times the capital intensity of your typical US industrial company...It’s clearly a worse business model than came before." I never underestimate the market's ability to ignore math in favor of speculative enthusiasm. But even if you can get beyond the upward price pressure the buildout is putting on electricity supply, DRAM, GPUs, etc. or the in-perpetuity costs of server and network depreciation, I still struggle with the pricing-power challenge. Tech giants have thrived on monopolies and duopolies. Now, they’re in an LLM arms race where it’s unclear when or even if ever leadership will be sustainable. The DoD pulling the plug on its relationship with Anthropic and moving over to a more ethically malleable partner in OpenAI has reinforced that concern. It's actually woven throughout my upcoming report on "Defense Innovation". Alongside "modernization", "sovereignty" is a guiding principal for accelerating defense spending. Europe is still engaged in debate about F-35 purchases due to concerns about whether or not the US has a "kill switch" for the jets. Will global governments really sign up to have foundation models built by and controlled by other countries woven into the fabric of their national security infrastructure? The sovereignty priority will likely give rise to "domestic AI champions" to service national security translating to more competition in the commercial arena. Xi Jinping said it and it's true across the globe: “We must have complete self-reliance, and comprehensively advance innovation, industrial development, and all AI-empowered applications, with mastery across all core AI technologies.” Learn more about Sage Road: https://lnkd.in/g6ZmVGmF. Interested in subscribing? Message me. Bloomberg link: https://lnkd.in/ghvCvxzj

10

California's RF/millimeter-wave startup, Movandi, has secured $40 million in funding to Accelerate Global 5G, Satellite, and Fixed Wireless Access Deployment. The round was led by a $20 million investment from ITHCA Group, the sovereign-backed technology investment firm driving innovation across the Middle East. Movandi’s RF chipsets, beamforming silicon, phased array antennas, and repeater systems have already transformed mmWave economics and performance. The company is leveraging these technologies to enable affordable FWA using 5G and Wi-Fi, non-terrestrial satellite communications, and advanced RF systems for high-throughput terrestrial and satellite networks. What do you think of Movandi's latest funding? ------------------------------------------------------------------------------------ ❓ Are you an RF engineer looking for your next opportunity? I'm offering 1:1 RF Job Search Consultations — focused on resume review, company targeting, and interview preparation. ✅ Book your session: https://lnkd.in/giENppTW 💡 Limited slots available — first come, first served.   

15

""This $100M funding propels TensorWave's mission to democratize access to cutting-edge AI compute," said Darrick Horton, CEO of TensorWave. "Our 8,192 Instinct MI325X GPU cluster marks just the beginning as we establish ourselves as the emerging AMD-powered leader in the rapidly expanding AI infrastructure market.” The new capital will fuel TensorWave’s operational growth, team expansion, and the accelerated deployment of its Instinct MI325X-powered training cluster. This growth comes at a pivotal moment, as demand for AI computing resources continues to outstrip supply and organizations seek alternatives to limited infrastructure options."

6

The AI Infrastructure Gigacycle: A Primer for 2026 My latest report via our new newsletter, The Dilligence Stack. The global semiconductor market is on track to reach $975 billion in 2027—within striking distance of the $1 trillion milestone originally anticipated for 2030. The sheer upside of this statistic seems only limited by fundamental constraints in the supply chain. This is an unprecedented moment in the semiconductor industry. One where countless conversations we have had with executives have yielded essentially the same comment/sentiment: “I’ve never seen anything like this in my 30-40-year career in this industry.” https://lnkd.in/gnXeE52i

11

Tower Semiconductor (TSEM) delivered impressive fourth-quarter 2025 results 📊🚀 (https://lnkd.in/g_3CTixt), reporting record revenue of $440 million -- up 11% sequentially and 14% year-over-year 📈 -- along with full-year revenue of $1.57 billion and a notable expansion in profitability, with Q4 net margins reaching 18%💰🔥. The standout driver of this performance is the company's silicon photonics (SiPho) platform 🌟, which has emerged as a critical enabler of AI-driven data center growth 🤖💻, positioning Tower as the majority supplier of 1.6T silicon photonic integrated circuits in partnership with leading hyperscalers and module manufacturers, including recent collaboration highlighted with NVIDIA 🤝⚡. Silicon photonics and silicon germanium revenues surged to 27% of the corporate total in 2025, with SiPho alone doubling to $228 million annually and achieving a Q4 run rate approaching $380 million 💥📈. In response to sustained and robust demand 📈🔥, Tower has committed an additional $270 million in capital expenditures — bringing the total investment to $920 million 💸 — to achieve more than a fivefold increase in silicon photonics capacity by the end of 2026 🏭🚀, with over 70% of this expanded capacity already reserved or in the process of reservation through 2028 via customer prepayments 🔒✨. The company has introduced an updated long-term financial model targeting $2.84 billion in annual revenue, approximately 40% gross margin, and $750 million in net profit (26% net margin) by 2028 📊🌟, reflecting a disciplined focus on value-driven growth and high-margin optical interconnect technologies that underpin the expanding infrastructure requirements of artificial intelligence 🌐🤖⚡.

34

1 Comment

Great to see SCINTIL Photonics Jim Theodoras presentation and the huge momentum behind CPO at PECC. Single wavelength CPO is a great first step but DWDM CPO enabled by #SHIP Heterogeneous integrated SiPho will be the ultimate transport architecture. The narrative is quickly shifting from "copper where you can, optics where you must" to a single unified DWDM CPO architecture in both #ScaleUp and #ScaleOut with power, reliability, cost, latency, reach and total bandwidth superior to copper .

26

KEY TAKEAWAY: Cadence of San Jose (USA) has built a $90 billion market cap company as the leading electronic design automation ("EDA") company for integrated circuits used in AI and hyperscale computing. Cadence's moat combines that of Lonza and other others who work with customers in designing the end product, with that of DNP Dai Nippon Printing Co., Ltd., whose photomask products require approval and compatibility both with chip designers and with fabs. ____________________________ NVIDIA uses Cadence tools in chip design, which are then manufactured at the TSMC fab. This creates a "triad", aka three sided moat, whereby its design tools are intertwined with NVIDIA's performance requirements, and TSMC's manufacturing process hardware, rendering switching costs much higher than they would be with only one of the two partners. For example, Cadence’s Reality Digital Twin platform creates detailed, physics-based virtual replicas of entire data centers to simulate and optimize their design and operation. By integrating Cadence’s Reality Digital Twin with NVIDIA Omniverse and their DGX/GB200 supercomputing systems, the partnership creates a unified, realistic simulation environment where AI factories can be designed, simulated, and optimized before being physically built. TSMC uses this to turn traditional 2D CAD designs into interactive 3D virtual models, including clean rooms, utility layouts, and manufacturing lines. So not only is Cadence very difficult to replace as a necessary piece in an extremely complicated jigsaw process, it is essential in designing the entire process. Cadence's net income and free cash flow had been growing 16%-18% annually for the decade prior to 2021. Since that time, net income growth has accelerated as its legacy EDA software tools continued to sell, but free cash flow has stagnated, as Cadence moved into the hardware business, and inventory has ballooned. For example, Cadence's Millennium M2000 is a hardware EDA platform, packed with NVIDIA Blackwell GPUs, and cutting edge networking hardware, designed to vastly outperform Cadence's legacy EDA software tools running on generic hardware. But customers have been slow to switch, since they require proof of the claimed productivity leap. This has created an inventory issue that Cadence did not have as a software vendor. It could be just a matter of timing, if the value is eventually proven. Remember...stocks are fun but always wear a helmet. As always this is not investment advice, nor an endorsement of any product nor service. Next up, endoscopy. Gianmarco Paolo Conti Clarke Jeffries Jason Celino, CFA Charles Shi Nay Soe Naing Joe Quatrochi, CFA Blair Abernethy Gary Mobley

1

Another key trend to watch in 2026 is the downstream server rack production ramp. According to NVIDIA's CEO comments two months ago, downstream appears to be a bigger bottleneck than upstream. Our “chip vs. rack” chart, which compares TSMC’s chip output with Blackwell server rack shipments, highlights quarterly fluctuations in TSMC’s production plan. As of November, monthly server rack output reached 5.5k units based on our Asia Tech Hardware analyst’s update (see Exhibit 8). If monthly output rises to 7k–8k in 2026, the annual run rate should reach 80k–90k NVL72 racks – enough to consume most of TSMC’s Blackwell chips.

1

Intel’s stock soaring on the news that Nvidia has taken a stake is one of those moments that makes the semiconductor story both fascinating and controversial. For decades, Intel and Nvidia were fierce rivals,competing on architectures, market share, and technological relevance. Now, in the face of shifting geopolitics, supply chain pressure, and escalating capital demands, we are watching the lines blur between competitor, collaborator, and shareholder. From my vantage point, this signals more than a financial investment. Nvidia is not in the business of symbolic gestures. A stake in Intel suggests a strategic calculation,that reinforcing U.S. domestic manufacturing capacity through Intel is not just good for the ecosystem, but ultimately good for Nvidia’s own long-term positioning. After all, AI workloads demand silicon at unprecedented scale, and no single company can control the entire chain from design to fabs to advanced packaging. The controversy here is whether this strengthens competition or quietly dilutes it. Can Nvidia remain an unbiased customer and ecosystem driver when it now has a vested interest in one of its historic rivals. Or is this a recognition that the capital and political stakes in semiconductors are now so high that even the most independent players must align for survival. For investors, this move also reframes how we think about value. It is not only about EPS or quarterly margins, but about national capacity, geopolitical resilience, and shared strategic bets. The irony is striking,competition once defined this sector, but collaboration may be the only way to secure its future. Is this the dawn of a stronger U.S. semiconductor ecosystem, or a signal that the traditional rules of rivalry no longer apply. #Semiconductors #Intel #Nvidia #AdvancedManufacturing #AIChips #TechnologyPolicy #StrategicInvesting #InnovationDebate

1

One implication of NVIDIA’s AI demand outlook that may be underappreciated: The bottleneck is shifting. As AI clusters scale toward tens—or even hundreds—of thousands of accelerators, the constraint is no longer just compute. It’s data movement. That’s driving rapid investment in optical interconnects. But the key issue isn’t simply adding capacity. It’s coordination. Across the stack: materials → epitaxy → foundry → packaging → system integration From what I’m seeing, the real constraints are increasingly: • yield and uniformity at wafer scale • packaging and assembly scalability • process portability across fabs • system-level integration In other words: The problem isn’t building optical devices. It’s manufacturing millions of identical ones, reliably and cost-effectively. That distinction matters. Because it shifts the investment question from: “Who has capacity?” to: “Who can coordinate the stack?” I wrote a deeper piece on this dynamic and where it may impact AI infrastructure: https://lnkd.in/e6HDB9KX #AIInfrastructure #Semiconductors #Photonics #Datacenters #ArtificialIntelligence #TechInvesting #OpticalNetworking 

5

2 Comments

The D-Wave / Quantum Circuits deal is another sign that the capital markets in quantum are finally maturing: public companies are no longer relying on single, monolithic architectures or vertically-integrated business models, but instead assembling best‑in‑class capabilities via acquisition to build useful offerings faster than pure R&D could. This mirrors what we’re seeing elsewhere: IonQ buying Oxford Ionics was a bold pivot (and a good one) to captialize and enable the latest in trapped‑ion scalability and semiconductor integration, Google aqui-hiring Atlantic Quantum's superconducting control stack, and now D‑Wave finally getting into gate-based systems and away from its questionable annealing approach. For investors, it is an exciting time and a great opportunity. The old model of “build everything in‑house” is giving way to a new, outward-facing-capital playbook: acquire and integrate specialized teams and IP that accelerates the path to commercial relevance (and revenue). We’re entering an era where the winners in quantum will be those that can seamlessly blend internal execution with external capability. D‑Wave announcement: https://lnkd.in/eQUmPKW3

51

1 Comment

🚀 Exit Fridays: Qualcomm's $2.4B Alphawave Bet - Building the AI Data Center of Tomorrow The Deal: In June 2025, Qualcomm announced its acquisition of UK-based Alphawave Semi for $2.4 billion in cash, marking a strategic push into AI-powered data centers. Alphawave shareholders will receive 183 pence per share—a stunning 96% premium over the company's March 31 closing price. Why It Matters: This isn't just another chip acquisition—it's Qualcomm's bold move to diversify beyond smartphones into the exploding AI infrastructure market. Alphawave's high-speed wired connectivity and compute technologies perfectly complement Qualcomm's power-efficient Oryon CPU and Hexagon NPU processors. The Strategic Fit: Alphawave specializes in chiplets—critical for enabling rapid data transfer in AI workloads. This gives Qualcomm access to full-stack AI deployments, from training and inference to networking and data storage across cloud and enterprise setups. Key Stats: 🎯 $21.9M in Q1 2025 royalties for Alphawave 🔥 Five new IP design wins in recent quarters 📈 Transaction expected to close Q1 2026 CEO Wisdom: "The combined teams share the goal of building advanced technology solutions and enabling next-level connected computing performance across a wide array of high-growth areas, including data center infrastructure." – Cristiano Amon, Qualcomm CEO Lessons for Startups: Complement, don't compete: Alphawave's connectivity IP filled Qualcomm's data center gaps perfectly Timing matters: The AI infrastructure boom created urgent demand for specialized solutions Premium valuations reward scarcity: 96% premiums reflect how critical these technologies have become The Future: This acquisition positions Qualcomm to compete directly with Broadcom and Marvell in custom silicon while building on its recent move into Nvidia's NVLink ecosystem. As one analyst noted, this deal gives Qualcomm "a running start in a part of the market that's expanding quickly thanks to cloud providers and hyperscalers". Will hardware connectivity become the new moat in AI infrastructure? #ExitFridays #Qualcomm #AIInfrastructure #SemiconductorM&A #DataCenter

1

1 Comment

Thrilled to announce our investment in @LightSpeed Photonics. Modern compute is constrained not just by silicon, but by how fast we can move data across a system. @lightspeedphotonics's near-packaged optical interconnects deliver high bandwidth at low power and plug straight into today’s boards, transferring data as light at up to 800 Gbps in a module small enough to balance on your fingertip. Building this means getting all of these to work right: micron-level alignment, thermal stability, low-outgassing materials, stray-light control, and manufacturable yields all have to work in harmony...and Rohin Y & Ramana V Pamidighantam have pulled it off. Looking forward to seeing them reach new heights! @piventures #deeptech

334

16 Comments