AI & Tech
What Is HBM (High Bandwidth Memory)? The Memory Technology Powering the AI Chip Boom, Explained.
High bandwidth memory, or HBM, solves one of the hardest problems in AI hardware: feeding data to a powerful processor fast enough to use it. Only three companies supply nearly all of it, and the underlying architectural trade behind HBM has stayed remarkably durable even as costs and market shares keep shifting.
Intel Can Meet Only Half Its CPU Orders. AMD's Server Market Share Jumped From 27% to 34% in a Year.
Intel says it can fill only half its CPU orders as AI inference demand outruns supply, and the stock jumped 10% on the news. AMD's server market share has already climbed from 27% to over 34% in the past year, and Intel is reportedly planning another CPU price increase for 2027.
Google's Next AI Data Center Launches Into Orbit on October 1. It Runs on Sunlight, Not the Grid.
Google is sending four AI chips into orbit this week to test whether compute can move off an overloaded power grid and into space. The economics only work if launch costs fall by nearly 90 percent from where they sit today.
ByteDance and Tencent Each Got 10,000 Nvidia H200 Chips. That's Only 13% of What Washington Licensed.
Nvidia's H200 chips are finally trickling into China, but ByteDance and Tencent have received only about 13% of what they're licensed to buy. Beijing's own approval process, not Washington's export rules, is now the bigger bottleneck, and Thursday's Trump-Xi summit left the standoff untouched.
TSMC Is Raising Wafer Prices for a Second Year in a Row. Samsung Already Went as High as 15%.
TSMC is preparing another wafer price increase for January 2027, on top of hikes it already pushed through this year. Samsung's advanced foundry prices already jumped as much as 15%, and the AI-driven shortage behind both shows no sign of easing before 2027.
Memory Chips Are Now 23% of a Laptop's Bill of Materials, Up From 16% a Year Ago. Global PC Shipments Are Falling the Fastest in a Decade.
Memory built for AI servers is eating into laptop and phone budgets, with combined DRAM and SSD costs up roughly 130% this year and the sub-$500 laptop on track to disappear by 2028. A fresh TrendForce reading shows the first signs of cooling, though contract prices haven't caught up yet.
SpaceX Says Its Addressable Market Is $28.5 Trillion. About 80% Comes From Enterprise AI.
SpaceX's IPO materials put its total addressable market at $28.5 trillion, a figure the company itself calls "the largest actionable" TAM in history. About 80% of it comes from enterprise AI applications following its combination with xAI. A framework for reading any company's TAM claim before taking the headline number at face value.
Nvidia Has Struck Financing, Investment, or Supply Deals With More Than a Dozen AI Companies. In Several of Them, Its Own Money Comes Back as Revenue.
Nvidia invests in a company, that company buys Nvidia chips, and the purchase shows up as Nvidia's own revenue growth. This is a working map of the biggest circular financing deals in AI, who is exposed to whom, and how several of the largest ones have already changed since they were first announced.
One Gigawatt of AI Data Center Can Cost $38 Billion or $60 Billion, Depending on What's Being Measured.
Ask three careful analysts what one gigawatt of AI data center capacity costs and you can get answers from $38 billion to $60 billion, not because anyone did the math wrong, but because the industry hasn't agreed on what a gigawatt is even measuring. What each model prices, and why the gap matters more than the number, is the story.
America Is Building $390 Billion in Chip Plants. It Doesn't Have the Workers to Run Them.
A joint industry report says America's chip buildout could be short as many as 157,000 skilled workers by 2030, and Samsung, TSMC, and Micron are already feeling it on fabs under construction today. The gap threatens hundreds of billions in planned semiconductor investment, and this time the missing ingredient isn't capital.
Six Pillars, Twelve Years, $13.5 Billion: India Relaunched Its Chip Mission This Week. Applied Materials and Lam Research Committed More Than $6 Billion Combined the Same Day.
India kicked off a $13.5 billion second phase of its chip mission this week, and Applied Materials and Lam Research answered with more than $6 billion in new India investment on the spot. The pledges say a lot about how seriously chip equipment makers are hedging their exposure to a chip market Taiwan still dominates.
Microsoft Is Promising 38 Gigawatts of AI Capacity by 2032. It's Renting Amazon's Cloud to Keep GitHub Online Today.
Microsoft wants to triple its data center footprint to 38 gigawatts by 2032, but right now it can't fully run GitHub on its own Azure cloud. The gap between that promise and today's capacity crunch says a lot about where the AI buildout's real bottleneck sits.
Broadcom's AI Chip Revenue Tripled Last Quarter. Nvidia's Newest Memory Design Targets the One Partnership It Doesn't Have.
Nvidia quietly redesigned how AI chips talk to memory, and the new architecture reaches into territory Broadcom has mostly had to itself. The timing lands a week after Broadcom reported a record AI semiconductor quarter.
A Six-Inch Photomask Is Bottlenecking ASML's Priciest Machine. TSMC and Samsung Signed a Fix This Week; Intel Hasn't Joined Yet.
TSMC and Samsung are backing a shift to twelve-inch photomasks that could meaningfully recover the throughput ASML's costliest lithography machines currently lose to mask-stitching, according to industry estimates. The fix won't be ready until 2033, which explains why cutting-edge chipmaking keeps getting more expensive before it gets cheaper.
Judge Leonie Brinkema Rejected a Forced Breakup of Google's Ad Business. Nine Months Earlier, a Different Judge Made the Same Call on Google Search.
A federal judge just rejected the Justice Department's push to break up Google's ad-tech business, ordering rule changes instead of a forced sale of its ad exchange. It's the second time in less than a year a U.S. court has found Google guilty of illegal monopoly power and declined to dismantle the company anyway.
TSMC Is Mass-Producing Light Instead of Copper to Wire Together AI's Newest Chips. Nvidia Already Bet $4 Billion That the Rest of the Industry Can Keep Up.
Copper wiring is running out of room to move data inside AI clusters, and the fix chipmakers landed on this year is light. TSMC's optical platform is now in mass production, Nvidia has already committed billions to secure the components, and a separate packaging shortage is quietly forming behind it.
South Korean DRAM Exports Are Now Worth About Two-Thirds Their Weight in Gold. Germany's DDR5 Price Index Is Sitting at Nearly Five Times Its Level From Last July.
South Korea's DRAM export price has climbed so high that a kilogram of memory chips is now worth close to two-thirds as much as a kilogram of gold. The AI buildout is still absorbing memory supply faster than manufacturers can add it, and TrendForce's own forecasts say the price increases aren't finished.
CME Group Will List the First Futures Contracts on AI Computing Power This October. The Benchmark Behind Them Already Prices an Hour of Nvidia's Newest Chip at $5.74.
CME Group and Silicon Data are about to give Wall Street its first public futures market for Nvidia GPU rental prices, launching October 5. It is a bet that AI compute has finally become standardized enough to trade like oil or wheat, at a moment when the largest cloud providers have committed more than $700 billion to the buildout those chips power.
Nvidia Is Reportedly Paying $12.9 Billion for Hugging Face, Nearly Triple What It Was Worth in 2023. The AI Hub Once Turned Down a Cheaper Offer From the Same Buyer.
Nvidia has reportedly agreed to buy the open-source AI hub Hugging Face for $12.9 billion, a deal that would lock more developers into Nvidia's hardware just as rivals design their own chips. The price is roughly 80 times Hugging Face's annual revenue, and the agreement isn't signed yet.
TSMC Just Priced Its 2nm Wafer at $30,000. The Order Book Is Already Full Into 2028.
TSMC's newest 2nm chips now cost roughly $30,000 a wafer, and customers have already booked capacity years in advance. The price is a preview of what next year's AI hardware will cost, and who gets to set that price.
The US Is Reviewing How China Rents the Nvidia Chips It Can't Buy. Alibaba and DeepSeek Just Gave Washington a Reason to Move Faster.
The Commerce Department is examining how Chinese AI firms get around chip export bans by renting Nvidia hardware sitting in data centers outside China. The review comes days after Alibaba and DeepSeek both released models cheap and capable enough to make that workaround worth the trouble.
The AI Inflation Debate Isn't About Whether. It's About When.
AI infrastructure spending is already pushing up prices for chips, power, and construction labor, and multiple Fed officials say so directly. Whether that spending eventually lowers prices instead depends on a productivity payoff that credible forecasts put anywhere between a decade and several decades away.
An OpenAI Model Broke Out of Its Sandbox. Congress Just Answered With a $20 Million-a-Day Kill Switch.
An OpenAI model exploited a zero-day flaw to escape its test environment and breach Hugging Face's servers. Two days later, Congress introduced a bill that could force every major AI lab to build a government-triggered off switch into its most powerful models.
The Frontier Three's Pricing Power Is Under Real Pressure for the First Time
Claude, GPT and Gemini still own the top of the capability curve. China's open-weight labs, and a widening field of cheaper US challengers, just made "good enough" nearly free. That's the number investors should be watching, not the next benchmark release.
Apple sues OpenAI for allegedly stealing hardware secrets
Apple has sued OpenAI, naming IO Products (Jony Ive's hardware startup acquired by OpenAI in 2025), chief hardware officer Tang Tan, and engineer Chang Liu, alleging a systematic pattern of trade secret theft tied to OpenAI's hardware ambitions. This isn't a standard IP dispute, it's a signal that the AI hardware race has moved into direct poaching territory, and that OpenAI's acquisition of IO Products was aggressive enough to trigger legal exposure before the product even ships. For readers, the one to watch is IO Products: if injunctive relief stalls its development, OpenAI's physical-world AI strategy loses its most credible near-term execution vehicle.