From the white hair to the billionaires, the smart guys who are emptied of the same frame are rich
Author:Beep News
Leopold created the Fund with $225 million, which was $5.5 billion over 12 months and has now expanded to $13 billionHIS BOTTLENECKS ARE ELECTRICITY, CALCULUS, MEMORY, LIGHT INTERCONNECTION, AND THE PHYSICAL INFRASTRUCTURE OF AI。
He didn't have an E.V.D. in his portfolio, but spent $8.46 billionWatching the dropDo the entire chip plate。
the white hair god who turned off his offer at a $6 share price, picked small capitalizations by the violet suaf theory, claiming 225 times the annualized earningsHis bottlenecks are the upstream of the AI-ray communications supply chain, such as the CPO light interconnection, the InP liner, and the light transceiver。
This theory was further emphasized in an interview with No Priors on 18 June 2026 by Intel CEO Chen Li Mu. Chen Liwu served as Cadence CEO for 12 years, before taking over Intel。
HE IS ALSO ONE OF THE MOST ACTIVE VENTURE CAPITALISTS IN THE SEMICONDUCTOR SECTOR, WITH INDIVIDUALS INVESTING IN MORE THAN 200 SEMICONDUCTOR COMPANIES, 159 OF WHICH ARE IPOSHis bottleneck covered new materials such as EDA, GaN/SiC/IP and light interconnection。
A CIRCUIT BOARD
PICK UP THE CIRCUIT BOARD OF ANY OF THE AI ACCELERATORS。
Before it is manufactured, designers need to validate the layout of hundreds of billions of transistor tubes with the EDA tool, replace silicon that is touching physical limits with new materials such as Imp, GaN, and SiC, and protect fine processes at every step of light and etching with helium。
ON BOARD, GPU CHIPS ANDHBMMemory stacked together, done by CoWOS or EMIB in IntelAdvanced EnvelopeI DON'T KNOW. GPU DECIDES THE UPPER LIMIT OF THE CALCULUS, HBM DECIDES WHETHER THE CALCULUS CAN BE RELEASED, AND THE SEAL DETERMINES WHETHER THEY CAN BE ASSEMBLED TOGETHER。
Between the plates and the plates, thousands of such accelerators work together. The copper cable is approaching the physical bandwidth limit, and the light interconnection is getting better。
Around the plate, 48V voltage is down to less than 1V required by GPU, and the heat is generated every step of the transition; the work of a 120kW cabinet, the traditional cold, is out of reachLiquid coldBeing a frame。
OUTSIDE, IT ALL NEEDS ELECTRICITY. AN AI DATA CENTRE USES ELECTRICITY EQUIVALENT TO A MEDIUM CITY, WHILE GRID EXPANSION AND NEW POWER GENERATION FACILITIES TAKE SEVERAL YEARS。
That's the whole picture of nine bottlenecks. It's broken down one by one。
Before the board
ALL CHIPS MUST BE DESIGNED AND VALIDATED BY THE EDA BEFORE THEY ARE MANUFACTURED, WITH CERTIFICATION LINKS ACCOUNTING FOR 60-70 PER CENT OF THE ENTIRE CHIP DEVELOPMENT CYCLE。
THE COMPLEXITY OF THE DESIGN CONTINUES TO EXPAND, BUT THE COMPUTATIONAL EFFICIENCY OF THE EDA TOOL IS NOT SYNCHRONIZED. THE COST OF FAILURE MAY EXCEED TENS OF MILLIONS OF UNITED STATES DOLLARS ONCE A PROBLEM HAS ARISEN THAT REQUIRES A REFLOW。
The EDA market was about $14.5 billion in 2025 and is expected to be close to $18 billion in 2026. The Symps, Cadence and Siemens took more than 65 per cent of the shares. Chen Li Moo has been CEO for 12 years in Cadence, and he knows better than most investors the pricing power of this chain, describing EDA as a gold mine. Cadence has been able to increase design contraction rates fivefold, and Siemens' AI system has achieved a 10-fold acceleration on some tasks。
Traditional silicon-based materials gradually touch the performance ceiling in the direction of power consumption, heat dispersion, light communication, etc. Five new materials are becoming breakthroughs: GaN (high frequency power), SiC (high voltage current), InP (light communication), man-made diamonds (thermal) and glass base plates (advanced seal)。
The 800G, 1.6T-ray modules are based on InP materials, and the current AI-light interconnection demand gap is approximately 40-60 per cent. The glass base plate is considered to be the next-generation advanced encapsulation direction, with both Intel and desk power increasing propulsion. Woleseed and Infineon invested more than $15 billion in SiC capacity only in 2025-2027。
At the beginning of 2026, an event that was completely ignored by most investors occurred: supply disturbances in Ras Laffan, Qatar, affected the global supply of 27-30 per cent of helium and short-term price increases of 40 to 100 per cent in spot prices。THE DEPENDENCE OF THE KOREAN SEMICONDUCTOR INDUSTRY ON QATAR HELIUM IS ABOUT 64.7 PER CENT, WHICH EXPOSES SAMSUNG AND SK HERCULES HBM TO SUPPLY RISKS。
Helium runs through EUV light, corrosive, sedimentary, crystal round cooling, non-renewable, with no alternatives. The semiconductor sector accounts for about 24 per cent of global helium consumption and is expected to increase to 30 per cent by 2030. More problematic is the increase in the consumption of helium per unit by about 20 per cent in the 2nm process compared to 3nm. The more advanced the process, the more dependent it is on a resource that is becoming scarce。
Samsung has introduced a helium recycling system with an advanced line recovery rate of 80-90 per cent for desk power. Recovery, however, can only be mitigated and the underlying problem cannot be resolved: supply is concentrated in a small number of sources of gas, and new sources are built on an annual basis。
On the board
HBM HAS PROVIDED HIGH-SPEED DATA TRANSMISSION CAPABILITIES FOR GPU AND HAS BEEN A CORE BOTTLENECK IN THE SUPPLY OF THE AIS SERVER。The memory ratio is nothing。
THE GLOBAL HBM MARKET SIZE IS PROJECTED TO BE AROUND $9.2 BILLION IN 2026 AND IS EXPECTED TO GROW TO NEARLY $70 BILLION BY 2035, WITH A COMPOUND INCREASE OF OVER 25 PER CENT PER YEAR. THE MARKET IS DOMINATED BY SK HERCULES, SAMSUNG AND MIRACLE, AND SK HERCULES, WITH LEADING PRODUCTION CAPABILITIES, HAS BECOME THE CORE SUPPLIER OF WEEDA, AND SAMSUNG AND MISAKI ARE ACCELERATING THE EXPANSION OF HBM3E AND HBM4。
GPU DECIDED ON THE UPPER LIMIT OF THE CALCULUS, AND HBM DECIDED WHETHER OR NOT THESE ALGORITHMS COULD BE RELEASED。
ADVANCED CONTAINMENT INTEGRATES GPU AND HBM INTO FULL AI ACCELERATOR, POWER BUILDCOWOSIT IS THE MOST MAINSTREAM PROGRAMME. EVEN WHEN GPU AND HBM HAVE BEEN PRODUCED, FAILURE TO COMPLETE THE ENCAPSULATION WILL NOT TRANSLATE INTO ARITHMETIC。
The CEO has publicly indicated that CoWOS is capable"Extremely nervous, sold out in 2026."I don't know. Production capacity has increased from about 35,000 to 40,000 tablets per month at the end of 2024 to the 2026 target of 120,000 to 140,000 tablets per month, but demand has increased more rapidly. The global demand for CoWos in 2026 is expected to be close to 1 million crystals, with Britain accounting for about 60 per cent of the total, and large volumes of capacity being targeted through long-term contracts。
Intel bets EMIB and the Glass Baseboard Program are attempting to compete with the build-up of power in the containment direction, and the ASE, Amkor and others are simultaneously expanding their production。
Between the plates
LARGE MODEL TRAINING REQUIRES THOUSANDS OR EVEN TENS OF THOUSANDS OF GPUS TO WORK TOGETHER. A SINGLE BLOCK OF GPU IS MORE POWERFUL AND THE ACTUAL UTILIZATION OF THE ENTIRE CLUSTER WILL BE DELAYED IF DATA TRANSMISSION BETWEEN CHIPS DOES NOT KEEP PACE. THE CURRENT MAINSTREAM COPPER CABLE INTERCONNECTION PROGRAMME IS APPROACHING THE PHYSICAL BANDWIDTH CEILING, WITH HIGH-SPEED INTERCONNECTIVITY CHIPS AND NEW INTERCONNECTIVE STRUCTURES BECOMING THE DIRECTION OF CAPITAL-INTENSIVE INPUTS。
PHOTOKOS IS THE NEXT GENERATION SOLUTION TO THE BOTTLENECKS OF INTERCONNECTION. SIGNAL DECAY AND HEAT PROBLEMS OCCUR IN TELECOMMUNICATIONS AT LONG DISTANCES AND IN HIGH-DENSITY TRANSMISSION SETTINGS, WHERE LIGHT SIGNALS HAVE PHYSICAL ADVANTAGES. SILICON PHOTONS AND CPOS (CO-ENCAPSULATED OPTICAL) ARE EXPECTED TO REDUCE THE USE OF INTERCONNECTIONS BY 30 TO 50 PER CENT, BUT MANUFACTURING PROCESSES, CONTAINMENT INTEGRATION AND COST CONTROL ARE NOT YET MATURE AND THERE IS A CLEAR GAP BETWEEN CAPACITY AND AI CLUSTER DEMAND. THE OPTICAL INTERCONNECTION MARKET WAS ABOUT $15 BILLION IN 2025 AND COULD REACH $43 BILLION BY 2034。
Wong In-hoon has invested almost all the companies that make light connections。Since 2026, NVIDIA has cumulatively invested more than $6.5 billion in photo-direction: approximately $2 billion each in Lumentum and Cohen, and $500 million in a layout silicon-photo route to Ayar Labs。
Around the board
The AI server requires 48Vs or even higher voltage and less than 1 Vs for multi-stage conversion to GPU operations. Traditional silicon-based instruments are inefficient in high-power settings and GaN and SiC are becoming next-generation programmes。
onsemi calculates that the power semiconductor value of the next generation of 1 MW AI racks has doubled from approximately $50,000 to $100,000. The GaN/SiC power plant market in 2025-2026 was about $2 billion and is expected to exceed $8 billion in 2030, with a compound rate of increase of more than 20 per cent per year。
Infineon's acquisition of GaN Systems to complete the product line, Navitas launched the GaN power programme for the AI data centre, and the expansion of SiC's capacity is being accelerated by onsemi, Woldese and STMicroelectronics。
A new generation of AI server cabinets, represented by NVIDIA GB200 NVL72, with more than 120 kW. If the heat is blown by a fan alone, the space and noise required for the room will go out of control. Liquid cooling is becoming the formula for the next generation of AI data centres。
THE GLOBAL DATA CENTRE LIQUID COLD MARKET WAS ABOUT $5 BILLION BY 2025 AND IS EXPECTED TO GROW TO $27.1 BILLION BY 2035. THE USE OF LIQUID COOLING IN THE NEW AI DATA CENTRE IS EXPECTED TO INCREASE FROM ABOUT 35 PER CENT IN 2025 TO ABOUT 55 PER CENT BY THE END OF 2026。
NVIDIA promotes liquid cold structures in Blackwell and Rubin platforms, and Microsoft, Google, Amazon and Meta accelerates their adoption in new data centres. In the area of chip-scale dissipation, Chen has set the direction of artificial diamonds, using its high-conductor heat capability to address the local concentration of high-power chips。
Outside the board
A large number of data centre projects in the United States have faced delays due to inadequate grid access。
Combined capital expenditure for Amazon, Microsoft, Google and Meta in 2026 is expected to reach $700 billion, with a significant share going to AI infrastructure and energy mix. Traditional grid expansion has not kept pace with demand and technology companies have begun to shift to alternative options such as long-term power purchase agreements, natural gas generation and nuclear power generation。
Leopold believes that there is a war behind Silicon Valley for all the remaining electricity contracts of this century and for every transformer. His judgment is that the real bottlenecks in the AI era are not algorithms, but electricity。
Williams invested $5.1 billion to build modular natural gas power facilities, and GE Vernova has a backlog of 100 GW-grade gas turbine orders; NVIDIA, through the NVentures, is investing in TerraPower to promote small modular nuclear reactors, and the Stargate project is exploring nuclear power supply。
Power construction involves grids, land, approval, longer construction cycles and more difficult to replicate quickly than other technological bottlenecks。
How long will this frame last
How long will this bottleneck investment framework take? It depends on when supply catch up with demand。
The second half of 2027 was the first supply release node in terms of the construction schedule: The SK Hercules M15X plant is scheduled to start production by mid-2027, and American Singapore and Taiwan plants are also targeted at 2027. White hair stock judgement supercycles will also start in mid-2027. The year 2028 was the second wave: the Samsung Pyeongtaek P5 factory, the SK Hercules American Indiana factory, and the U.S. Hiroshima factory。Chen's judgment is, "No relief until 2028."
But new capacity on the line does not mean that bottlenecks disappear. Each generation of GPU demand for HBM is doubling, and the demand for HBM4 from the next generation of the NVIDIA Rubin structure will be further amplified; and super-large cloud manufacturers have locked up large amounts of new capacity through long-term contracts, with limited market openings。
BETWEEN 2017 AND 2018, THE PRICE OF DRAM SURGED, WITH A SUBSTANTIAL EXPANSION OF SAMSUNG AND A MORE THAN 50 PER CENT INCREASE IN CAPITAL EXPENDITURE. AFTER THE RELEASE OF THE NEW PRODUCT IN 2019, PRICES COLLAPSED AND THE ENTIRE INDUSTRY LOST. IT TOOK 18 MONTHS TO REVERSE ENERGY INPUTS TO PRICES。
THIS ROUND IS MUCH LARGER THAN THE PREVIOUS ROUND. DRAM PRICES ARE PROJECTED TO INCREASE BY ABOUT 275-300 PER CENT FROM 2025 TO 2027, THREE TIMES AS HIGH AS IN 2017-2018 AND AT A THREE-FOLD BASE. THE MARKET VALUE OF THE SK HERCULES, SAMSUNG AND U.S.S.S. MEMORY MANUFACTURERS HAS EXCEEDED $1 TRILLION, AND THE HBM PROFIT MARGIN IS AS HIGH AS 60-70 PER CENT, FAR ABOVE TRADITIONAL DRAM. IF THE SAME 18-MONTH WINDOW WERE TO BE APPLIED, THE PERIOD BETWEEN THE END OF 2028 AND MID-2029 WOULD BE A PERIOD OF HEIGHTENED VIGILANCE。
THERE IS A REAL NEED TO FOCUS ON THE SIGNAL THAT IF CAPITAL SPENDING IN AI SLOWS DOWN BY THAT TIME, WHILE THE NEW CAPACITY OF THE THREE PRODUCERS IS RELEASED, SUPPLY-DEMAND RELATIONSHIPS CAN QUICKLY REVERSE, BOTTLENECKS BECOME EXCESSIVE AND PRICING RIGHTS RETURN FROM SUPPLIERS TO BUYERS。
Leopold's operation suggests he's already preparing for the scene. While making multiple power and infrastructure, he used $8.46 billion to make empty semiconductor plates。HIS JUDGEMENT WAS THAT, ONCE THE AI INFRASTRUCTURE CONSTRUCTION CYCLE PEAKED, FIERCE COMPETITION AMONG CHIP COMPANIES WOULD REDUCE PROFITABILITY, BUT THAT THE SCARCITY OF ELECTRICITY AND PHYSICAL INFRASTRUCTURE WAS MORE PERSISTENT AND DIFFICULT TO REPLICATE。
Until then, the supply and demand imbalances in the chain had not been alleviated。
