The AI Chip War ▍Topic hub
From NVIDIA's GPUs to hyperscalers' custom ASICs, and the HBM and CoWoS bottlenecks behind them — the AI compute value chain is being redrawn. The chip war powering the AI era.
What Is a Custom AI Chip (ASIC), and How Does It Differ From NVIDIA's GPUs?
A custom AI chip, or ASIC, is hardwired for one specific task, trading NVIDIA GPU-style flexibility for lower power use and cost per operation. Google, Amazon, Meta, and Microsoft now build ASICs to cut costs and reduce supplier dependence, but long design cycles and low flexibility mean GPUs still lead on new, fast-changing models.
Why DDR5, PCs, and Phones Are All Getting More Expensive: Inside the Memory Supercycle
Memory makers have redirected production toward AI servers and HBM, squeezing consumer DRAM and NAND supply and pushing Q1 2026 contract prices up 90–95% and retail DDR5 prices up roughly 500% in twelve months. Dell, Lenovo, and Samsung Electronics have already raised prices, Xiaomi has warned of 2026 phone price increases, and forecasts show the increases only slowing — not reversing — through 2026, with meaningful new capacity not arriving until 2027.
Google's $120 Billion Marvell Deal, Explained: How Custom AI Chips Are Eating Into NVIDIA's Lead
The widely reported '$120 billion' Google-Marvell deal is not an order or a revenue figure — it is a cumulative custom-chip purchase threshold Marvell must hit by fiscal 2033 to fully vest warrants Google holds. Marvell's own guidance that this Google revenue won't ramp meaningfully until fiscal 2029 triggered an 8%-plus stock drop, even as Amazon, Meta, and Google all scale in-house AI silicon against an NVIDIA that still commands roughly 70% of the AI chip market.
CoWoS Capacity, Not GPU Wafers, Is the Real Limit on AI Chip Supply
TSMC's CoWoS advanced packaging, not logic wafer output, is the real constraint on AI GPU shipments: capacity stays booked solid through 2026 even as TSMC doubles output annually and pushes its 2022–2027 CAGR above 80%, while overflow orders worth roughly 180,000–190,000 wafers a year go to Amkor.
CoWoS, Not Logic Wafers, Is the Real Constraint on AI Chip Supply
TSMC's CoWoS advanced packaging, not logic wafer output, is the binding constraint on AI accelerator supply through 2025–2026. TSMC discloses only growth rates while TrendForce estimates monthly capacity near 120,000–140,000 wafers by end-2026, and TSMC has begun outsourcing overflow to Amkor and ASE Technology because a 18–24 month expansion cycle keeps packaging capacity from catching up with GPU demand.
The AI Power Ladder: From a 1,200-Watt GPU to a Gigawatt Data Center
A single Blackwell B200 GPU draws up to 1,200 watts, and packing 72 of them into an NVIDIA GB200 NVL72 rack pushes power density to roughly 120 kW — more than ten times a traditional rack's 5–10 kW. xAI's Colossus cluster already pulls an estimated 250 MW across about 100,000 H100 GPUs, while industry PUE averages 1.54 against Google's 1.09. Cloud operators now list electricity supply as a primary constraint on where new data centers can be built.
DRAM Prices Nearly Doubled in a Quarter: Why This Memory Supercycle Is Different
General-purpose DRAM contract prices jumped roughly 93–98% quarter-on-quarter in Q1 2026, with NAND Flash close behind. The surge stems from memory makers diverting capacity to HBM and AI servers under long-term agreements, leaving standard DRAM supply unable to catch up. Micron's 2026 HBM output is already sold out through 2027, and record revenue at SK Hynix and Micron shows the demand is structural, not a typical inventory-driven cycle.
What Is an Advanced Process Node? What the "nm" in 2nm Actually Means
The "2nm" in a modern process node is not a literal measurement — node names decoupled from physical transistor size after roughly the 22nm generation. "2nm" instead signals a manufacturing generation built on Gate-All-Around transistors, with TSMC's N2, Samsung's SF2, and Intel's 18A all reaching mass production in Q4 2025.
What Is HBM, and Why Is It the Bottleneck Choking AI Chips?
High Bandwidth Memory (HBM) stacks DRAM dies vertically with through-silicon vias to break the memory-wall bottleneck, and its newest generation, HBM4, reaches 2 TB/s per stack. But supply sits with only three makers — SK Hynix at 58%, Samsung Electronics and Micron at 21% each — so HBM output and advanced packaging capacity, not GPU logic alone, now set the pace of AI chip shipments.
TSMC Adds NT$10, TAIEX Jumps 356 Points to Close Above 46,300
Taiwan's benchmark TAIEX closed at 46,331.45 points on August 28, up 356.23 points and above the 46,000 threshold, on turnover of roughly NT$1.023 trillion. TSMC rose NT$10 to NT$2,420, and 36 listed stocks hit their daily limit after NVIDIA's earnings beat lifted sentiment at the open.
DRAM's Supercycle: Why This AI-Driven Memory Shortage Breaks the Old Pattern
DRAM contract prices jumped 93-98% quarter-on-quarter in Q1 2026, with a further 58-63% rise forecast for Q2, as capacity shifts toward HBM and AI servers. NAND Flash contract prices are set to climb 70-75% QoQ in the same quarter. SK Hynix and Micron posted record revenue growth, and HBM supply for 2026 is fully sold out through 2027, marking a structural, AI-driven cycle rather than a traditional supply-side one.
NVIDIA Guides to 70% Revenue Growth for 2027 as Q3 Outlook Tops $108 Billion and Stock Reverses to a 4.12% After-Hours Gain
NVIDIA's CFO guided 2027 revenue growth to 70%, above a prior 45% market estimate, while Q2 revenue hit $96.2 billion (up 106% year over year) and Q3 guidance reached $108 billion, both beating analyst consensus. Data Center revenue rose 117% to reach 92% of total sales, supply purchase commitments doubled to $279 billion, and gross margin held at 75% before a guided step-down. Shares closed down 1.59% but reversed to $218.3, up 4.12%, after hours.
NVIDIA Reportedly Agrees to Acquire Hugging Face for $12.9 Billion, Its Largest Buyout Ever
NVIDIA has agreed to acquire Hugging Face for approximately $12.9 billion, a deal first disclosed by The Information, that would rank as NVIDIA's largest-ever corporate acquisition — surpassing its $6.9 billion Mellanox purchase in 2020. Neither company has confirmed the transaction, and no signed agreement has been reported. The price caps a rapid climb in Hugging Face's valuation, from $4.5 billion in 2023 to a reported asking price above $13 billion weeks before the disclosure.
Nvidia Reportedly Preparing Employee-Funded PAC 'NVPAC' as Executives' $443,000 RNC Donation Surfaces
Nvidia is reportedly preparing to launch an employee-funded political action committee named NVPAC to expand its influence over U.S. policymaking, joining Meta Platforms and Alphabet's Google in forming such vehicles. The move comes as Federal Election Commission records show two Nvidia executives, Gavin Sherry and Sundeep Madra, donated $443,000 to the Republican National Committee on June 24. Nvidia has declined to comment on the reports.
NVIDIA Q2 FY2027 Results: Revenue Hits $96.2B, Up 106% as Data Center Tops $89B
NVIDIA reported Q2 FY2027 revenue of $96.2 billion, up 106% year-over-year and above the roughly $92 billion consensus, with Data Center revenue of $89.0 billion, up 117% and about 92% of total sales. Gross margin held at 75.0% and non-GAAP EPS of $2.22 beat the $2.10 estimate. Q3 guidance of $108.0 billion excludes any China Data Center compute revenue.
Why AI GPUs Are Scarce and Expensive: Unpacking the HBM and CoWoS Bottlenecks
AI GPUs stay scarce and expensive because production runs through several constrained stages at once, not one single choke point. HBM supply is concentrated among SK Hynix, Samsung Electronics, and Micron; CoWoS advanced packaging capacity sits with TSMC (台積電); and vertical integration across chip, memory, packaging, and software gives suppliers pricing power. Any one tight stage caps shipments regardless of raw chip fabrication capacity.
Inside the AI Data Center Chip War: NVIDIA GPUs, Cloud ASICs, and the Inference Startups
AI data centers draw on three chip camps: NVIDIA's general-purpose GPUs, cloud giants' self-designed ASICs such as Google's TPU and AWS's Trainium/Inferentia, and startup inference chips from Groq, Cerebras, and SambaNova. NVIDIA leads on its CUDA ecosystem and versatility, while cloud providers build their own silicon to cut costs and reduce dependence on outside suppliers.
What Is Model Quantization? How It Shrinks Large AI Models to Fit Your Computer
Model quantization compresses neural network weights from high-precision formats like 16-bit floating point into low-precision formats such as 8-bit or 4-bit integers, cutting a model's memory footprint to roughly a quarter of its original size and letting models that once needed data-center hardware run on ordinary computers instead.
What Is CoWoS Advanced Packaging, and Why Does It Give TSMC Its Grip on AI Chip Supply?
CoWoS is TSMC's 2.5D packaging technology that places GPUs and HBM memory on a shared silicon interposer. It has become the real limit on AI chip supply because interposer yield and process complexity, not wafer fabrication, set the ceiling on how many accelerators can ship.
What Is a Sovereign Wealth Fund? Why They're Piling Into NVIDIA and TSMC
Sovereign wealth funds—state-run investment vehicles built on oil revenue or foreign-exchange surpluses—have made NVIDIA and TSMC core long-term holdings during the AI wave. Norway's NBIM, the world's largest at roughly $2 trillion, discloses a $62 billion NVIDIA stake (1.28% of the company) and a $34 billion TSMC stake, but most other sovereign funds disclose far less, making NBIM's numbers a floor rather than the full picture.
What Are Advanced Process Nodes? What TSMC's 2nm and 3nm Numbers Actually Measure
Chip "process node" labels like 2nm and 3nm no longer describe a physical measurement — they mark a manufacturing generation defined by transistor density and efficiency. The real contest is over performance-power-area gains, a shift from FinFET to Gate-All-Around transistors, and dependence on a single EUV equipment supplier.
AI Accelerator Chips Explained: How GPUs, TPUs, and ASICs Differ
AI accelerator chips split into three types — GPUs, TPUs, and ASICs — each trading generality for efficiency differently. NVIDIA's GPUs lead on CUDA's flexibility, while TPUs and inference-focused ASICs like Groq's LPU and AWS Inferentia sacrifice versatility for efficiency, with HBM versus on-chip SRAM memory determining which model sizes and latency needs each chip serves best.
What Is MoE (Mixture of Experts)? Why It Makes Large Language Models Cheaper and Faster
Mixture of Experts (MoE) architecture separates a large language model's total parameter count from its per-token activated parameter count, letting models such as Qwen3 235B-A22B run inference at roughly the speed of a much smaller dense model while keeping a large model's knowledge capacity — the trade-off is that memory requirements still scale with total parameters, not activated ones.
Why Does AI Need HBM? How High Bandwidth Memory Breaks the Memory Wall
HBM solves AI's core bottleneck—the memory wall—by stacking DRAM on a wide silicon-interposer bus beside the GPU, lifting bandwidth from 256 GB/s (HBM2) to 2 TB/s (HBM4) and capacity from 8GB to 64GB, though supply stays concentrated in three makers led by SK Hynix's 58% share.
SpaceX and NVIDIA Plan to Launch Orbital AI Data Centers Starting Next Year, Musk Says
SpaceX plans to launch its first NVIDIA-chip-powered AI satellites, Starmind AI1, in the fourth quarter of 2027, reaching large-scale deployment by 2028 — at least a year ahead of the original schedule. The company has also filed with the FCC for a network of up to 1 million satellites, while Taiwanese suppliers Unitech and Sesoda report rising satellite-related revenue and order shares tied to the buildout.
NVIDIA Launches Jetson Orin Nano 2 Robotics Computer, Targets Entry-Level Edge AI
NVIDIA announced the Jetson Orin Nano 2 robotics computer on August 25, 2026, packing 78 TOPS of AI compute, 8GB of memory, and an 8-core Arm CPU while doubling inference performance and cutting power draw 40% versus its predecessor. Modules and developer kits ship in the first half of 2027, targeting a robotics developer base NVIDIA says already tops 3 million.
NVIDIA's Groq 3 LPX Inference Rack Enters Full Production, Nebius First to Deploy
NVIDIA confirmed on August 24, 2026 that its 256-chip Groq 3 LPX inference rack has entered full production, with Nebius named as the first cloud provider set to deploy it later this year alongside Vera CPU and Rubin GPU.
OpenAI's 700W Jalapeño Chip Claims Up to 1.9x Efficiency Over Nvidia's GB300 in First Published Benchmarks
OpenAI says its Broadcom-built Jalapeño chip delivered 1.5x-1.9x more throughput per kilowatt and 1.7x-3.6x lower latency than Nvidia's GB200/GB300 on SemiAnalysis's InferenceX suite, though the lead narrows to about 1.5x under utility-power and multi-token-prediction comparisons, and Nvidia's upcoming Vera Rubin platform was not tested.
AI Agents Are Reshaping More Than GPUs: How NVIDIA's Vera Rubin Rebuilds the Entire AI Factory
NVIDIA's Vera Rubin platform moves agentic AI infrastructure into production across three layers at once: SpaceX and SpaceXAI are deploying the Vera CPU, Nebius is adopting the newly mass-produced Groq 3 LPX inference system, and CoreWeave is running Spectrum-X Multiplane networking that scales to 512,000 GPUs — turning the AI factory into one coordinated, full-stack system rather than a single-chip upgrade.
NVIDIA's August 26 Earnings: A Record $92 Billion Forecast Meets a Post-Report Selloff Pattern
NVIDIA reports fiscal Q2 earnings on August 26, with analysts projecting record revenue of $92 billion, above the $78 billion initial-year estimate. After 14 straight quarters of beating forecasts, analysts say NVIDIA needs 95% profit growth and $51.5 billion in net income to clear the bar again, while options price a 5.3% post-earnings swing and its shares have fallen after each of the last four reports.
NVIDIA to Raise AI Server Prices by More Than 15% Starting Early Next Year as Memory Costs Surge
NVIDIA has told its largest customers that AI server prices will rise by more than 15% on Grace Blackwell and Vera Rubin systems shipping early next year, as memory suppliers led by Samsung, SK Hynix and Micron struggle to keep pace with demand and NVIDIA opts to pass rising memory costs to customers rather than absorb them within its roughly 75% non-GAAP gross margin.
Broadcom in Talks for Over $60 Billion in AI Chip Debt Financing, Bloomberg Reports
Broadcom is negotiating with a group of lenders to raise more than $60 billion in debt financing for AI chip infrastructure benefiting Anthropic and other companies, with the package potentially reaching $100 billion once a proposed $30 billion junior-debt tranche and a $60-70 billion Broadcom-guaranteed senior-secured tranche are combined. Blackstone and Apollo Global Management are among the institutions in talks to participate.
NVIDIA Takes Minority Stake in Data Center Developer Cloverleaf Infrastructure
NVIDIA announced a minority equity investment in Cloverleaf Infrastructure to jointly accelerate development of large-scale U.S. computing facilities, targeting land, power, cooling, and surrounding infrastructure constraints. Founded in 2024, Cloverleaf already runs multiple gigawatt-scale projects and plans to adopt NVIDIA's DSX platform for integrated site planning from the earliest development stages.
Google's Marvell Deal Rattles MediaTek Shares, but Foreign Investors Call It Overreaction
Google's up-to-$12.2-billion equity deal with Marvell briefly rattled MediaTek shares, which fell more than 4% intraday and 12.7% over four sessions on AMD/Marvell AI-ASIC rumors. J.P. Morgan and Morgan Stanley say the deal expands Marvell's role around Google's TPU ecosystem rather than replacing core TPU work or MediaTek's existing orders. Foreign institutional investors maintain a Buy rating on MediaTek, citing a multi-vendor strategy and a NT$6,800 target.
MediaTek Shares Drop Over 4% as Google Signs Up to $12.2 Billion Stock-Warrant Deal With Marvell
MediaTek shares fell more than 4% on August 20, 2026 after Google signed a warrant agreement giving it rights to purchase up to $12.2 billion in Marvell shares, a deal that could generate up to $120 billion in Marvell revenue through fiscal 2033. Marvell shares jumped 9.85% while Broadcom fell 4.57% in the same session, as Google added a third ASIC partner alongside Broadcom and MediaTek.
NVIDIA Partners Bring Physical AI and Smart Automation to Automation Taipei 2026
NVIDIA's partner network — including Advantech, FANUC, Solomon, Techman Robot, ITRI, Delta, Universal Robots, and Spingence — is showing physical AI and edge robotics at Automation Taipei 2026, held at Taipei Nangang Exhibition Center Halls 1 and 2. Jetson Thor, Jetson AGX Thor, Jetson Orin, Isaac Sim, Isaac Lab, and Omniverse are powering humanoid robots, AMRs, quadrupeds, and digital-twin factories across the show floor.
NVIDIA's AI Moat Shifts From Chips to Capital as Huang Backs OpenAI With Up to $105 Billion
NVIDIA (輝達) has pledged up to $105 billion in funding for OpenAI's Ohio data center and is separately pursuing a $500 billion chip-financing agreement with Wall Street firms. CNBC frames the moves as NVIDIA's AI moat shifting from chips to capital, a reading backed by NVIDIA's own numbers: quarterly free cash flow up 18-fold to $48.5 billion and marketable equity holdings up from $12.9 billion to $30.2 billion.
Malaysia's HBM Shipments Climb as Intel's Advanced Packaging Reaches for TSMC's CoWoS Crown
South Korean High Bandwidth Memory (HBM) exports to Malaysia have reached roughly $1.3 billion while shipments to Taiwan fell below $3 billion, down from over $5 billion months earlier. The shift coincides with Intel's newly completed Pelican advanced-packaging plant in Malaysia and reports that TSMC (台積電) has begun routing overflow CoWoS orders to Intel's Malaysian facility.
TrendForce Trims 2026 Global TV Shipment Outlook as Memory Prices Squeeze Cost Structure
TrendForce cut its 2026 global TV shipment forecast to a 0.6% annual decline, totaling 194.81 million units, after 2025 shipments already fell 0.8% to roughly 196.20 million units. Even as DRAM's share of TV bill-of-materials cost jumped from 2.5-3% to 6-7% on a more-than-fourfold rise in DDR4 contract prices, TrendForce raised its Mini LED TV penetration forecast to 10%, with TCL positioned to exceed 30% share in that segment.
NVIDIA and OpenAI Expand AI Mega Data Center Push to 10 GW Power Scale
OpenAI, SB Energy, and NVIDIA are building an 8 IT-GW AI campus in Ohio, backed by NVIDIA's $1.5 billion SB Energy investment and a prior $100 billion commitment to OpenAI, while SB Energy and SoftBank plan at least 10 GW of new power generation and $4.2 billion in grid infrastructure to support it.
NVIDIA Cuts OpenAI Data Center Financing Guarantee From a Discussed $250 Billion to Sub-$120 Billion First Phase
NVIDIA has restructured its financing backing for OpenAI's Ohio data center campus, with the originally discussed $250 billion guarantee narrowed to below $120 billion for the first phase and finalized at $105 billion for a 20-year lease. The revised deal covers roughly 5GW of an eventual 8GW-to-10GW site; funding for the remaining 5GW is still undecided.
Trump Administration Discourages Apple From Using CXMT Chips as Micron, SK Hynix, SanDisk Shares Surge
U.S. Commerce Secretary Howard Lutnick said the Trump administration has told Apple that using Chinese memory chips is not advisable, a stance he confirmed applies to Changxin Memory Technologies (CXMT), which already holds a 7% global DRAM revenue share. On the same day, Micron rose more than 5.9%, SK Hynix rose 5.6%, and SanDisk closed up 8.9%.
NVIDIA's Spectrum-X Photonic Switch Enters Mass Production, Built With TSMC, SPIL, Foxconn
NVIDIA announced its Spectrum-X Ethernet co-packaged-optics (CPO) switch has entered full mass production, calling it the industry's first 200G/lane CPO Ethernet switch system<CITE:E1>. The hardware is manufactured through a Taiwan-anchored supply chain spanning TSMC, SPIL, Lumentum, TFC and Foxconn<CITE:E5>.
NVIDIA in Talks to Invest $3 Billion in SB Energy to Build OpenAI's Ohio Data Center
NVIDIA is negotiating a direct investment of up to $3 billion in SoftBank's SB Energy, split into two $1.5 billion tranches, to help build an Ohio data center for OpenAI, while the three parties also discuss up to $100 billion in credit support for the project.
US Commerce Secretary Confirms Washington Told Apple Not to Buy Chinese Memory Chips
US Commerce Secretary Howard Lutnick says the Trump administration explicitly told Apple (蘋果) it does not want the company sourcing memory chips from China, even as Apple tests China's CXMT and YMTC chips amid an AI-driven shortage, and faces pressure from state senators and rival chipmaker Micron.
Indonesia Launches First University AI Center: UGM, Indosat and NVIDIA Open NVAITC in Yogyakarta
According to NVIDIA's blog, Indonesia's Ministry of Communication and Digital Affairs, Indosat Ooredoo Hutchison, NVIDIA and Universitas Gadjah Mada (UGM) launched the UGM Indosat NVIDIA AI Technology Center (NVAITC) in Yogyakarta this week — the country's first university-based AI technology center. Powered by NVIDIA's full-stack AI platform and Indosat's GPU Merdeka platform, NVAITC already backs projects tied to Indonesia's more than 1 million annual tuberculosis cases and a farm sector employing nearly 30% of the workforce.
Not TSMC: NVIDIA's Second-Largest Equity Stake Is Musk's SpaceX
According to a Liberty Times (LTN) report citing NVIDIA's SEC filing, NVIDIA held about 128.8 million Class A shares of SpaceX as of the end of the second quarter, making it SpaceX's sixth-largest investor. CNYES reported the stake was valued at roughly $21 billion at quarter-end — a position trailing only NVIDIA's Intel holding among its disclosed equity stakes.
UBS: Micron's Standard DRAM Margins Set to Overtake HBM, Nearing 90%
According to a UBS report cited by TechNews and Cnyes, Micron's (美光) standard DRAM gross margin is projected to climb from roughly 44%-50% in 2025 to 89%-91% by mid-2026 and 92% by year-end, overtaking HBM margins of 56%-70% over the same period. UBS attributes the reversal to HBM's heavier wafer and packaging demands squeezing standard-memory supply, plus simultaneous AI-server demand for both HBM and DDR5.
Norway's $2.3 Trillion Wealth Fund Discloses First-Ever SpaceX Stake, Alongside TSMC and NVIDIA Holdings
According to TechNews and a CNA report carried by UDN Money, Norway's $2.3 trillion sovereign wealth fund disclosed for the first time on August 12, 2026 that it holds a 0.05% stake in SpaceX, alongside 1.7% of TSMC (台積電, worth $34 billion) and 1.28% of NVIDIA (輝達, worth $62 billion), based on holdings as of June 30, 2026.
NVIDIA's Nemotron 4: A Reported Push to Rival the World's Top Open-Source AI Models
According to Taiwan's Central News Agency (CNA), citing The Information, NVIDIA (輝達) is developing a new model family called Nemotron 4 to compete with the world's top open-source AI models. The largest version is expected to carry at least 1 trillion parameters — about twice today's flagship Nemotron 3 Ultra, per Inside.com.tw — with training still incomplete and no confirmed launch date, though staff cite late fall as a possible target.