Weekly: Micron leans into memory supercycle
10 min read.
Highlights
Memory supercycle. Bloomberg does a great, long piece on the memory crunch. It seems like every week there’s a deep dive into the current memory supercycle. Contracts are moving from an annual to a quarterly basis as the power shifts from the client to the supplier. Consumer electronics are facing a tight squeeze as AI hoovers up all the supply.
In light of this, WSJ reports that Micron is investing US$200 billion to build out its memory manufacturing capacity. This is a radical departure from historical memory business tactics, which are known to be cyclical; memory chipmakers tend to ride out the ups and downs of the industry. Micron’s moves are a bet that the increased demand is a structural shift in the memory industry, or at the very least that the AI demand will stay for the foreseeable future.
Amazon’s AI capex. The FT delves into Amazon’s US$200 billion capital expenditure plans to boost its AI capacities, including its in-house Trainium chips. CEO Andy Jassy consolidated chip, AI model, and research teams into a single unit, cutting costs and jobs and streamlining Amazon’s AI functions.
India’s chip ambitions. The Economist covers why India’s data centre industry is booming, which I covered in today’s Daily newsletter. Over the past week, during India’s AI summit, there were several announcements, including governmental ambitions to move chipmaking to India and several deals between Nvidia, AMD, and local tech companies.
Thanks for reading.
Table of Contents
Debby Wu, Takashi Mochizuki, and Yoolim Lee, “Rampant AI Demand for Memory Is Fueling a Growing Chip Crisis,” Bloomberg, 02/16/2026.
Robbie Whelan, “Micron Is Spending $200 Billion to Break the AI Memory Bottleneck,” WSJ, 02/16/2026.
Rafe Rosner-Uddin, “Amazon’s Andy Jassy bets on $200bn AI spending drive to revive AWS,” FT, 02/14/2026.
The Economist, “India is in the midst of a data-centre investment boom,” The Economist, 02/19/2026.
Dylan Patel, Cam Quilici, Bryan Shan, et al., “InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX,” SemiAnalysis, 02/17/2026.
1.
Debby Wu, Takashi Mochizuki, and Yoolim Lee, “Rampant AI Demand for Memory Is Fueling a Growing Chip Crisis,” Bloomberg, 02/16/2026.
A growing procession of tech industry leaders including Elon Musk and Tim Cook are warning about a global crisis in the making: A shortage of memory chips is beginning to hammer profits, derail corporate plans and inflate price tags on everything from laptops and smartphones to automobiles and data centers — and the crunch is only going to get worse.
Since the start of 2026, Tesla Inc., Apple Inc. and a dozen other major corporations have signaled that the shortage of DRAM, or dynamic random access memory — the fundamental building block of almost all technology — will constrain production. Cook warned it will compress iPhone margins. Micron Technology Inc. called the bottleneck “unprecedented.” Musk got to the intractable nature of the problem when he declared Tesla is going to have to build its own memory fabrication plant.
The fundamental reason for the squeeze is the buildout of AI data centers. Companies like Alphabet Inc. and OpenAI are gobbling up an increasing share of memory chip production — by buying millions of Nvidia Corp. AI accelerators that come with huge allotments of memory — to run their chatbots and other applications. That’s left consumer electronics producers fighting over a dwindling supply of chips from the likes of Samsung Electronics Co. and Micron.
The resulting price spikes are starting to look a bit like the Weimar Republic’s hyperinflation. The cost of one type of DRAM soared 75% from December to January, accelerating price hikes throughout the holiday quarter. A growing number of retailers and middlemen are changing their prices every day. “RAMmageddon” is the term some use to describe what’s coming.
What’s worrying about the trend is that prices are soaring and supplies are running dry even before the AI giants really get going with their data center construction plans. Alphabet and Amazon.com Inc. just announced plans for a construction blitz this year that could reach $185 billion and $200 billion, respectively, more money than any company in history has poured into capital expenditures in a single year.
A manager at a laptop maker said Samsung Electronics has recently begun reviewing its memory supply contracts every quarter or so, versus generally on an annual basis. Chinese smartphone makers including Xiaomi Corp., Oppo and Shenzhen Transsion Holdings Co. are trimming shipment targets for 2026, with Oppo cutting its forecast by as much as 20%, Chinese media outlet Jiemian reported. The companies did not respond to requests for comment.
2.
Robbie Whelan, “Micron Is Spending $200 Billion to Break the AI Memory Bottleneck,” WSJ, 02/16/2026.
Micron Technology is the largest American maker of memory chips—the tiny slices of silicon that store and transfer data and help power everything from smartphones and car computers to laptops and data centers. Micron is rushing to add manufacturing capacity to avert the biggest supply crunch the memory industry has seen in more than 40 years.
In Boise, where the company is based, Micron is spending $50 billion to more than double the size of its 450-acre campus, including the construction of two new chip factories, or fabs. The first fab’s inaugural silicon wafers are expected to roll off the factory line in mid-2027, making DRAM, a type of memory used to make the high-bandwidth memory chips, or HBM, that are increasingly essential to advanced artificial-intelligence computing. Both plants should be in production by the end of 2028.
Each fab will be 600,000 square feet—the size of more than 10 football fields—making them some of the biggest “clean rooms” ever built in America. To prepare the site, engineers have already blasted through more than 7 million pounds of dynamite. An army of construction workers, building contractors and architects have set up a small city’s worth of trailers so they can work around the clock.
Each Boise fab is expected to use 70,000 tons of steel (almost as much used to build the Golden Gate Bridge) and 300,000 cubic yards of concrete (enough for four Empire State Buildings).
That’s not all. Near Syracuse, Micron just broke ground on a $100 billion fab complex that represents the state of New York’s largest-ever private investment. Late last year, Micron announced a $9.6 billion fab investment in Hiroshima, Japan, while competitor SK Hynix announced in January that it would build a $13 billion fab in South Korea, in addition to a $4 billion manufacturing complex it is building in Indiana.
3.
Rafe Rosner-Uddin, “Amazon’s Andy Jassy bets on $200bn AI spending drive to revive AWS,” FT, 02/14/2026.
Amazon is embarking on the largest capital spending programme in its history, seeking to regain momentum against AI rivals by expanding data centres, developing chips and building models.
The group is undergoing a strategic shake-up amid fears its cloud arm, AWS, is losing ground to competitors in securing corporate AI contracts, according to more than a dozen current and former senior employees.
Chief executive Andy Jassy last week announced Amazon’s capital expenditure would rise to $200bn this year, exceeding that of Google and Microsoft, with the outlay focused on computing infrastructure.
In December, he consolidated the group’s chip, model and advanced research teams under a single leadership structure, a move intended to align its AI plans. The Amazon chief has also cut costs, including jobs — eliminating some 30,000 of about 350,000 corporate roles.
It has touted the uptake of the group’s Graviton and Trainium chips, used for conventional cloud computing and AI training respectively. Sales of these chips are on course to generate more than $10bn in combined annual revenue.
Amazon debuted its latest generation of Trainium chips in December, promising a significant increase in performance. It is holding talks to join OpenAI’s latest multibillion-dollar funding round in a move partly designed to ensure the ChatGPT maker adopts its semiconductors, said people familiar with the matter.
The chips should also help Amazon reduce its reliance on Nvidia’s products, helping to expand AWS’s profit margins from renting out data-centre capacity to corporate customers.
4.
The Economist, “India is in the midst of a data-centre investment boom,” The Economist, 02/19/2026.
Tour around Navi Mumbai, the younger and less glamorous eastern sibling of India’s financial capital, and among the chemical plants, oil refineries and industrial parks you will spot a handful of gleaming new buildings. Few people enter through their carefully monitored security gates, and little traffic passes by outside. These vast sheds are data centres, the infrastructure intended to underpin India’s goal of becoming an artificial-intelligence superpower.
Data centres are rapidly appearing across the country. Its installed capacity reached 1.3 gigawatts (GW) last year, according to JLL, a property firm. That may be small compared with America (38.7GW) or China (9.5GW), but is nearly triple the figure from 2020. And the growth is set to continue. During a global AI summit held in Delhi this week, the Adani Group, an Indian conglomerate, announced it would pour $100bn into data centres by 2035. Others such as NTT DATA, a Japanese IT giant that is currently the largest owner of data centres in India, also have big investment plans, as do America’s hyperscalers. India will struggle to rival America and China in the development of cutting-edge AI. But it is fast becoming an important hub for data centres.
That is down to three things. The first is a government push to localise data. In 2024 India, with its vast population, produced by some estimates around a fifth of the world’s digital information—but had only 3% of its data-centre capacity. Since 2018 the Reserve Bank of India has mandated that financial institutions retain clients’ data within the country. A law set to come into force next year may require that certain kinds of personal data, including some of the many social-media messages sent in the country, also remain there.
Whether Indians will benefit from all this is unclear. Data centres themselves only generate a significant number of jobs while they are being built. Jensen Huang of Nvidia, the leading seller of AI chips, has argued that the data-centre boom could be as good for India’s economy as the internet. That is optimistic. India’s data centres, once erected, become ghostly presences. Still, the country can reassure itself that it will soon control much more of the plumbing through which its enormous volume of data passes.
5.
Dylan Patel, Cam Quilici, Bryan Shan, et al., “InferenceX v2: NVIDIA Blackwell Vs AMD vs Hopper - Formerly InferenceMAX,” SemiAnalysis, 02/17/2026.
InferenceXv2 (formerly InferenceMAX) builds on the foundation established by InferenceMAXv1, our open-source, continuously updated inference benchmark that has set a new standard for AI inference performance and economics. InferenceMAXv1 moved beyond static, point-in-time benchmarks by running continuous tests across hundreds of chips and popular open-source frameworks.
InferenceXv2 builds on this foundation. It expands coverage to include large scale DeepSeek MoE disaggregated inference (disagg prefill, or simply “disagg”) with wide expert parallelism (wideEP) optimization to all 6 NVIDIA western GPU SKUs from the past 4 years as well as to every single AMD western GPU SKU released in the past 3 years – in total InferenceXv2 utilizes close to 1000 frontier GPUs for a full benchmark run across all SKUs.
With today’s release, InferenceXv2 is now the first suite to benchmark the Blackwell Ultra GB300 NVL72 and B300 across the whole pareto frontier curve, and it is the first third party benchmark to test disagg+wideEP multi-node FP4 and FP8 MI355X performance. In future iterations of InferenceX, we will continue to focus heavily on disaggregated serving with wide expert parallelism as that is what is deployed in production at Frontier AI Labs like OpenAI, Anthropic, xAI, Google Deepmind, DeepSeek as well as advanced API providers like TogetherAI, Baseten, and Fireworks. In this article, we will also break down the system engineering principles and economics in play around the latest Claude Code Fast mode feature.
Our benchmark is completely open-source under Apache 2.0 – this means that we are able to move at the same rapid speed at which the AI software ecosystem is advancing. If you like our work and would like to show us some support, please drop a star on our GitHub! We also provide a free data visualizer at https://inferencex.com for everyone in the ML community to explore the complete dataset themselves.
–