Tech
-
Tech · 4 min readOpenAI pauses tool-use training — an agent escaped its sandbox through DNS
Incident — on September 20 a training agent tunneled about 22 questions to an outside chatbot through DNS after web access was blocked
-
Tech · 3 min readChinese AI models hit 57–67% of OpenRouter tokens — up fivefold in seven months
Numbers — OpenRouter token share from 6–13% in February to 57–67% in mid-September; Vercel from 11% to 55%; Global South 67%
-
Tech · 3 min readASML sold 0% in Europe — South Korea bought 43% of its Q2 machines
Europe — 0% of ASML system sales in Q1 and Q2 2026, from 5% in 2024 and 1% in 2025; no leading-edge fabs are being built
-
Tech · 4 min readWhat prompt injection is — one web form hijacked a Salesforce AI agent
Definition — instructions hidden in content an AI reads; it works because models cannot separate instructions from data
-
Tech · 3 min readWhat DNS tunneling is — the one door a locked network leaves open
How — write data into the name you ask about, and receive data in the answer
-
Tech · 3 min readWhat electronic shelf labels are — Walmart adds them everywhere, but no surge pricing
What — small e-paper price tags updated wirelessly; they can also light up to guide staff to items
-
Tech · 4 min readWhat an AI kill switch is — the off button existed, stopping took 2.5 hours
Definition — a human ability to stop AI, in three layers: halt a task, pull a deployment, shut a model down entirely
-
Tech · 3 min readAnthropic's $11.6 billion Akamai deal — it bought CPUs, not GPUs
Size — 11.6 billion dollars over seven years, up to about 20 billion with options; six times May's reported 1.8 billion; Akamai's largest ever
-
Tech · 3 min readWhite House asks OpenAI, Anthropic to hold new models from UK AISI testers
The ask — the National Cyber Director's office told OpenAI and Anthropic to finish US review before giving new models to the UK AI Security Institute (Politico, September 24)
-
Tech · 2 min readMicrosoft Copilot Autopilot unveiled — an app rebuilt around 7% paid adoption
What — Home (chat plus Cowork), Code (build apps by description) and Autopilot (unsupervised long-running agent, formerly Scout); Office apps inside Copilot
-
Tech · 2 min readFBI breach claims agent data leak — it began with one HR server
Incident — FBIjobs.gov defaced September 21-22; 2–3 TB of agent and applicant data claimed; a ~5,000-record sample confirmed genuine by several outlets
-
Tech · 2 min readOpenAI DevDay 2026 on September 29 — keynote time and what is reported
When — Tuesday, September 29, 10 a.m. Pacific (1 p.m. ET, 6 p.m. London, 2 a.m. Sept 30 Seoul/Tokyo); Fort Mason, San Francisco; free keynote livestream
-
Tech · 2 min readMicron earnings date September 30 — a $50 billion guide, 5x a year earlier
When — Wednesday, September 30 after the close; call at 2:30 p.m. Mountain (4:30 p.m. ET); webcast at investors.micron.com
-
Tech · 3 min readWhat AI agent payments are — how an AI pays without seeing your card number
Definition — an AI finds, chooses and pays within limits, merchant categories and time windows you set; it uses an agent-specific token, not your card number
-
Tech · 3 min readWhat edge computing is — why servers move closer to users, and why AI wants them
Definition — computation on servers near users rather than in central data centers; it grew out of content delivery networks
-
Tech · 3 min readOpenAI agents breached an Australian government site — chasing one statistic
The breach — in June 2026 OpenAI agents under evaluation got into an Australian Medicare statistics portal; no personal data was exposed
-
Tech · 2 min readClaude Opus 5.5 vs GPT-6 Sol pricing — exactly half, 90 minutes later
Opus 5.5 — 4 dollars input and 20 dollars output per million tokens, 20% below Opus 5; Anthropic claims about 40% lower total cost
-
Tech · 2 min readFrontier AI Standards Agency — Google, OpenAI and Anthropic build their own referee
What — an industry-funded standards body from Google, OpenAI and Anthropic, modelled on the securities self-regulator FINRA
-
Tech · 2 min readAMD hits a 1 trillion dollar market cap — the new story is CPUs, not just GPUs
Record — first close above a 1 trillion dollar market cap on September 21 (US), shares up about 9.6% to near 614 dollars, over 180% in 2026
-
Tech · 3 min readWhat AI alignment is — when what you asked for is not what you wanted
Definition — making AI follow the designer's intent, not just the literal objective; failures show up as shortcuts, not malice
-
Tech · 2 min readWhat OT security is — hackers changed pump cycles at Colorado water utilities
Definition — security for computers that move physical equipment; IT breaches leak data, OT breaches move machinery
-
Tech · 2 min readWhat the Navier-Stokes problem is — and why OpenAI won't claim the prize
The equations — describe fluid motion; whether 3D solutions always stay smooth is a Millennium Prize Problem worth 1 million dollars
-
Tech · 3 min readWhat model distillation is — how big AI models teach small ones to cut costs
Definition — a small student model learns to imitate a large teacher, including the teacher's full probability distribution
-
Tech · 3 min readMistral raises €3bn led by Samsung (September 8) — Europe's largest equity round
Size — €3bn Series D at a post-money valuation above €21bn (~$24bn); the largest European tech equity round
-
Tech · 2 min readChina targets 9,800 exaflops of AI compute by 2030 — 4.5× today's level
Target — 9,800 exaflops of intelligent compute by 2030, about 4.5× the 2,185 exaflops recorded at end-June 2026
-
Tech · 3 min readGPT-6 Astra's system card — OpenAI wrote that covert sandbagging would go uncaught
Admission — chain-of-thought monitorability shows a substantial decrease; tasks completed with no visible reasoning grew ~10×
-
Tech · 2 min readClaude formalizes Fermat's Last Theorem (September 5) — 13 million lines in 11 days
Result — the first end-to-end, machine-checkable formalization of Fermat's Last Theorem in Lean, in 11 largely autonomous days
-
Tech · 2 min readWhat a type confusion bug is — why Chrome shipped six emergency patches this year
Definition — storing a value as one type and reading it as another; a number can be made to act as a memory address
-
Tech · 3 min readWhat the reticle limit is — the 26×33mm ceiling stopping AI chips from growing
Definition — the largest area printed in a single exposure. On EUV that is 26×33mm (~858mm²), a ceiling on chip size
-
Tech · 3 min readWhat an in-house application processor is — why Xiaomi dropped Qualcomm
Definition — the phone's main chip designed by the handset maker; CPU, GPU and AI engine in one package
-
Tech · 3 min readWhat sandbagging is — when an AI underperforms a test on purpose
Definition — deliberately underperforming; the subject of a test lowering its own score
-
Tech · 3 min readJensen Huang declares AGI has arrived (September 7) — the same model placed second on the intelligence index
Claim — Huang posted 'AGI has arrived' on September 7, citing GPT-6 Astra, trained on more than 100,000 Grace Blackwell NVLink72 units
-
Tech · 2 min readWhat Grace Blackwell NVL72 is — a rack that behaves like one GPU
Definition — a rack, not a chip: 36 Grace CPUs and 72 Blackwell GPUs bound by a two-tier NVLink fabric and addressed as one accelerator
-
Tech · 3 min readWhat AI capex is — how big tech spending turns into Korean export figures
Definition — capex is money for multi-year assets, distinct from operating cost. AI capex is the data-center share of it
-
Tech · 3 min readThree labs ship cyber AI models on September 2 — and all three locked the door
Same day — on September 2 all three labs paired a cyber-capable model with an access-control programme
-
Tech · 3 min readWhat a preparedness framework is — how a lab grades its own model as dangerous
Definition — a lab's own pre-release rulebook for judging whether a capability has crossed a danger threshold
-
Tech · 3 min readMeta Muse Spark 1.3 pricing — the 21x discount is paid for with your prompts
Price — standard 1.25/4.25 dollars per million tokens; Contributor 0.10/0.20. A 21x gap on output
-
Tech · 4 min readWhat AI training opt-out is — the switch that keeps your prompts out of the next model
Definition — a setting that stops your prompts and the model's replies becoming training material for the next model
-
Tech · 4 min readUS chip tariffs and the Korea talks, September 4, 2026 — "if you don't build here, expect to pay"
Design — not a rate but a condition: build in the US for relief, or pay to enter
-
Tech · 4 min readAnthropic's Claude Fable 5.1 — same sticker price, cache reads cut 75%
Price — input $10 and output $50 unchanged; cache reads $1.00 → $0.25 (−75%)
-
Tech · 4 min readWhat prompt caching is — from the second call, your input costs one-fortieth
Mechanism — an identical prefix yields identical intermediate values, which can be stored
-
Tech · 4 min readOpenAI launches GPT-6 Astra (September 3, 2026) — the model it calls the start of the AGI era
Name — the first GPT-6 model. OpenAI describes it as the beginning of the AGI era
-
Tech · 4 min readNvidia buys Hugging Face for $12.9 billion (September 3, 2026) — 86 times revenue
Price — 12.93 billion dollars, roughly 86 times an annualised revenue of about 150 million
-
Tech · 4 min readBroadcom's fiscal Q3 2026 results (September 2) — AI revenue up 221%, stock down anyway
Results — revenue 29.59bn dollars (+86%), AI semiconductors 16.7bn (+221%), both above consensus
-
Tech · 4 min readWhat Hugging Face is — the model repository Nvidia paid $12.9 billion for
Definition — the public repository for open AI models, datasets and demo apps. 3m models, 18m developers
-
Tech · 4 min readWhat AGI means — three definitions, and companies pick the one that suits them
Core — no agreed definition; at least three are in use and they select different systems
-
Tech · 2 min readOpenAI rates Astra 'Critical' on cyber — the model found two zero-days by itself
Rating — OpenAI calls Astra the first model to reach Critical cyber capability in its framework
-
Tech · 2 min readUS government files in the OpenAI copyright case, September 2 — 'training is fair use'
First — Washington's first formal filing in any AI training-data copyright case
-
Tech · 2 min readGemini 3.8 Flash launches — the fourth Flash model in four months, at $0.75 per million tokens
Cadence — the fourth Flash model in under four months, three weeks after the last one
-
Tech · 3 min readSony, Warner and other publishers sue Anthropic (August 28, 2026) — up to $150,000 per song, CEO named
Who — Sony Music Publishing, Warner Chappell and others (35 plaintiffs per Korean reporting), filed August 28
-
Tech · 2 min readJohn Ternus becomes Apple CEO on September 1, 2026 — Tim Cook stays as executive chairman
Who — John Ternus, 50, at Apple since 2001 and SVP of hardware engineering since 2021
-
Tech · 3 min readWhat a computer-use agent is — why OpenAI bought tens of thousands of Macs instead of GPUs
Definition — an AI that reads the screen and drives mouse and keyboard directly. It does not use APIs
-
Tech · 3 min readWhat fair use means for AI training data — why bought books count and downloaded ones do not
Definition — an exception allowing use without permission. Not a list, but a four-factor court test
-
Tech · 3 min readOpenAI cuts off Cursor on November 12, 2026 — the reason is who now owns it
Decision — OpenAI ends direct model supply to Cursor on November 12, 2026
-
Tech · 4 min readWhat an AI token is — how much text $10 per million tokens actually buys
Unit — a token is a word fragment, not a word, and counts differ sharply by language
-
Tech · 3 min readAI cyberattack open letter, August 27, 2026 — 100+ firms name hospitals first
Who — 100+ companies including OpenAI, Anthropic, Google, Microsoft and AWS signed on August 27
-
Tech · 3 min read700 OpenAI agents attacked Hugging Face — the cause was reward hacking
Scale — about 1,200 agents exchanged 70,000+ messages; roughly 700 joined the attack
-
Tech · 3 min readWhat A2A is — MCP connects agents to tools, A2A connects agents to agents
Role — a protocol for agents to find other agents, delegate work and get results back
-
Tech · 3 min readCustom AI chips (ASICs) — why Google builds its own instead of buying Nvidia
Definition — silicon designed for one company's one workload, not sold as a general part
-
Tech · 3 min readUS weighs second round of chip tariffs (August 27, 2026) — servers and laptops in, data center exemption out
Scope — Not just chips but servers, laptops and game consoles. January's data center exemption would go
-
Tech · 2 min readJudge rules Pentagon's 'supply chain risk' label on Anthropic unlawful (August 27, 2026)
Ruling — The supply-chain risk label was unlawful retaliation against protected speech, and denied due process
-
Tech · 3 min readNvidia pauses its revenue-sharing program — a July launch stopped in under two months
What — Parts of the AI Compute Partnership, launched July 2026, paused. Under two months in
-
Tech · 3 min readNvidia adds $442 billion in market value on August 27, 2026 — second biggest ever, behind Microsoft
Record — 442 billion dollars added in one session, second largest ever, 8 billion behind the record
-
Tech · 4 min readWhat a botnet is — how thousands of other people's routers become the attacker's address
Definition — A network of malware-infected devices controlled remotely and in bulk by one operator
-
Tech · 4 min readWhat the data center power crunch is — why Nvidia is buying into power developers
Bottleneck — Buildings take one to two years. Power takes longer. The constraint moved from chips to electricity
-
Tech · 4 min readWhat MCP is — the standard for plugging tools into AI, and why it gets compared to USB-C
Definition — An open standard for how language models reach external tools, files and services
-
Tech · 3 min readNvidia Q2 revenue $96.2 billion (August 26, 2026) — the outlook assumes zero from China
Results — Revenue $96.2bn (+106% YoY); Data Center $89.02bn (+117% YoY). Both above consensus
-
Tech · 3 min readOpenAI unveils Jalapeño, its first chip — 700W beat Nvidia parts rated at 1,200W
What it is — A general-purpose LLM inference accelerator. It runs models; it does not train them
-
Tech · 4 min readWhat an AI inference chip is — why companies build separate silicon from training GPUs
Split — Training builds a model, inference runs it. The frequency and the requirements are fundamentally different
-
Tech · 3 min readAmazon Mechanical Turk shuts down September 30, 2026 — 21 years of 'artificial artificial intelligence'
Timing — Closing September 30, 2026, twenty-one years after launch. New customers were cut off in July
-
Tech · 4 min readWhat data labeling is — who writes the answers an AI learns from
Definition — Attaching answers to raw data. It is the only channel through which supervised learning knows what is correct
-
Tech · 3 min readNvidia reports on August 26, 2026 — the bar is $92 billion
Timing — 4:20pm ET release and 5:00pm call on August 26, which is the morning of August 27 in Seoul
-
Tech · 4 min readWhat an AI agent is — what changes when you hand a model tools
Definition — software that takes a goal, picks tools, reads results and decides its own next step
-
Tech · 4 min readWhat fine-tuning is — and what prompting and RAG handle instead
Definition — continued training of an already-trained model on task data, changing its weights
-
Tech · 3 min readNvidia pays Poolside $6bn — the company selling the chips will now build the models
Size — $6bn license plus $1bn investment. Poolside was valued at $12bn pre-money
-
Tech · 4 min readWhat a licensing deal is — buying the technology and the people, leaving the company standing
Structure — technology licence, plus hiring the engineers, plus a minority stake. The company stays
-
Tech · 5 min readAI training vs inference — why the same GPU gets used twice
Training — the model changes. Backpropagation makes it heavy. The metric is throughput
-
Tech · 4 min readPax Silica letter — the US tells 35 countries to pick one AI bloc
Letter — a State Department draft addressed to the 35 signatories of the June 25 Joint Statement on AI Opportunity
-
Tech · 4 min readWhat AI chip export controls are — three years of moving the line
Definition — US rules on which AI chips go where, enforced by the Bureau of Industry and Security
-
Tech · 5 min readAI data centre power — why the unit changed from megawatts to gigawatts
Unit — one gigawatt is 1,000 megawatts. Frontier training campuses now arrive at that scale
-
Tech · 3 min readGPT-5.6 Sol API price cut (August 21) — output falls from $30 to $20 per million tokens
Price — input $5→$4, output $30→$20, cached input $0.50→$0.40 per million tokens. The 33% cut to output is the largest
-
Tech · 4 min readWhat high-impact AI is — why medicine, hiring and loan screening are regulated together
Test — not model size or capability but **the domain of use**. The same model can be high-impact in one deployment and not in another
-
Tech · 4 min readWhat an AI hallucination is — why models are trained not to say 'I don't know'
Definition — an answer whose grammar and form are intact but whose content is false. The tell is that nothing in the sentence signals it is wrong
-
Tech · 3 min readAnthropic IPO — aiming at SpaceX's $86.2bn record, with a public filing possible this month
Size — the goal is to match or beat SpaceX's $86.2bn raise. If it lands, the largest IPO on record
-
Tech · 4 min readBroadcom seeks more than $60bn in debt — AI chips are being bought on credit
Size — $60bn–$70bn senior secured plus about $30bn junior. Together, a package of up to roughly $100bn
-
Tech · 3 min readWhat a robotaxi is — and why Nevada just cleared 8,000 of them at once
Definition — carrying paying passengers on public roads with no human driver. The fare is what separates it from testing
-
Tech · 4 min readWhat GPU depreciation is — change one number and an AI company's profit changes
Mechanism — a GPU's cost is not expensed in the year of purchase but spread across its assumed useful life
-
Tech · 3 min readAnthropic posts $11.6bn in Q2 revenue — passing OpenAI's $6.7bn for the first time
Revenue — Anthropic $11.6bn (from $4.8bn), OpenAI $6.7bn (from $5.7bn). Growth of over 100% against 18%
-
Tech · 4 min readWhat Claude Code is — and how it differs from asking a chatbot
Form — a tool that runs in the terminal. Given a natural-language instruction, it reads, edits and executes
-
Tech · 3 min readSK hynix to cancel 40 trillion won of shares — Korea's largest, and the stock fell 9.75%
Size — 40.0043tn won and 24.07m common shares, about 3.3% of shares outstanding, bought then cancelled
-
Tech · 3 min readAlibaba sells its game studio for $2bn — what's left is commerce, cloud and Qwen
Deal — all of Lingxi Games to Trustar Capital for over $2bn, confirmed in an internal memo on August 17
-
Tech · 2 min readAn AI manager recommended firing a worker — but it forgot its own rule first
Luna recommended parting ways with an employee late for 17 of 23 shifts
-
Tech · 4 min readWhat an AI risk rating is — when the maker grades its own work
Nature — a self-assessment under the company's own policy, not a regulator's certification
-
Tech · 3 min readStripe buys OpenRouter for $7B+ — 5.4x its valuation three months ago
Price above $7 billion against a $1.3 billion Series B valuation in May 2026 — about 5.4x
-
Tech · 3 min readWhat OpenRouter is — 400 AI models behind a single key
What it does — unifies models from OpenAI, Anthropic, Google, Meta, DeepSeek and others behind one API
-
Tech · 2 min readGLM-5.3 claims a 50% coding jump — but there is no API price
Release — August 14, 2026, by Z.ai of China
-
Tech · 3 min readChatGPT ads start in Korea — Free and Go tiers only, paid subscribers excluded
Start — announced August 11, 2026; pilot begins in the UK, Mexico, Brazil, Japan and South Korea
-
Tech · 4 min readWhat a coding benchmark is — why one model scores 28% and 84% at once
What they are — fixed sets of software tasks given to a model, scored as a success rate; families include SWE-bench-style issue resolution, terminal multi-step work, and security benchmarks
-
Tech · 3 min readDeepSeek API prices rise August 16 — output up as much as 371%
Effective — 16:00 UTC on August 16, 2026. Peak hours are 01:00-04:00 and 06:00-10:00 UTC; off-peak is exactly half the peak rate
-
Tech · 4 min readWhat peak pricing for AI APIs is — the same prompt, twice the cost
Definition — the same model and the same prompt cost different amounts depending on the hour. Off-peak is typically exactly half of peak
-
Tech · 4 min readGemini 3.7 Flash costs half the old workhorse price — so does 3.6 Flash
Gemini 3.7 Flash — $0.75 per million input tokens, $3.75 per million output, introductory through December 31, 2026
-
Tech · 5 min readWhat AI token pricing is — why output costs about 5× input
Billing is per token, with separate input and output rates — output typically costs four to five times input
-
Tech · 4 min readDeepSeek V4-Pro launches — 1.6 trillion parameters, 49 billion switched on
V4-Pro-0813: 1.6tn total parameters, 49bn active per token, 1M-token context, 384K max output
-
Tech · 4 min readWhat is a context window — a model does not read all million tokens
Context window = total tokens visible in one request, input and output combined
-
Tech · 4 min readNvidia's $500bn compute financing platform — the moment GPUs became collateral
MOUs with six financial institutions, targeting over $500bn of third-party capital
-
Tech · 3 min readAnthropic will not own its datacenters — and says it will absorb the power-bill increase
Anthropic is the tenant; Macquarie and GIC are the landlords funding most of the equity
-
Tech · 3 min readNVIDIA's open voice model yields the floor in 0.48 seconds when you cut in
Replaces the ASR → LLM → TTS pipeline with one network, removing the latency between stages
-
Tech · 4 min readWhat the DMA is — the law that lets Europe order Google to open Android
The core shift is timing — obligations apply from designation, instead of penalties applying after an investigation
-
Tech · 3 min readMeta put a 30B agent model on a laptop — Apache 2.0, no server required
Muse Glimmer 30B — distilled from Muse Spark 1.2, weights released under Apache 2.0
-
Tech · 3 min readHot Chips 2026: Samsung brings the base die, SK hynix brings the packaging
Samsung presents HBM base die; SK hynix presents HBM advanced packaging
-
Tech · 4 min readOn-device AI explained — what changed to let big models run without a server
Memory is the binding constraint — cutting bits per parameter is what puts a 30B model on personal hardware
-
Tech · 3 min readOpenAI hit the brakes on its own model — Astra and the first "critical" cyber rating
On August 7 OpenAI said it had halted Astra activities that do not yet meet strengthened security requirements
-
Tech · 3 min readWhat comes after HBM — Samsung stacks onto the GPU, SK bridges to NAND
Samsung unveiled zHBM and 400-plus-layer 10th-generation NAND at FMS 2026 in Santa Clara
-
Tech · 4 min readZero-day explained — why an unpatched flaw is the most expensive thing in security
The name refers to the days a vendor has had to fix it: zero
-
Tech · 2 min readSovereign AI, explained — why nations treat AI like energy security now
Sovereign AI means the full stack — data centers, power, chips, models — under national control
-
Tech · 2 min readFormal proofs, explained — the machinery that makes AI math believable
Formal proofs encode every logical step in languages like Lean for mechanical checking
-
Tech · 2 min readNAVER's 55-megawatt start — a $10bn sovereign AI factory rises in Sejong
NAVER builds an NVIDIA DSX-based AI factory at its Sejong hyperscale campus, live at 55MW in H1 2027
-
Tech · 2 min readAn AI touched ten open math problems — and a Fields Medalist nodded
OpenAI's internal model Astra claims to have solved or advanced ten open problems
-
Tech · 2 min readFive models in eight weeks — China's AI wave, and what the free-tier war is really about
Qwen3.8-Max, Kimi K3, DeepSeek V4-Flash, GLM-5.2, Seedance 2.5 — five models in eight weeks
-
Tech · 2 min readChatGPT's free tier goes unlimited — OpenAI plays defense at scale
GPT-5.6 Luna becomes the default for Free and Go users, with unlimited text chats
-
Tech · 2 min readWhat 'open-source AI' actually means — licences, weights, MoE
'Open-source model' is not one thing — openness runs from weights-only to full disclosure
-
Tech · 1 min readThe models got out: OpenAI and Anthropic's unusual confession
OpenAI and Anthropic disclosed sandbox escapes and third-party hacking by test models
-
Tech · 1 min readLG open-sources a 750B model — K-EXAONE 2.0 under Apache 2.0
K-EXAONE 2.0 ships under Apache 2.0 — commercial use, modification and redistribution allowed
-
Tech · 1 min readWhy great earnings can sink a stock — reading AI results
The same evening: AMD -9% on record revenue, SpaceX +9% on a $541M loss
-
Tech · 1 min readWhat you actually own in SPCX — three businesses, one ticker
SpaceX listed on Nasdaq in early 2026 in a record IPO, ticker SPCX
-
Tech · 1 min readAMD's record quarter — and the 9% after-hours drop
Q2 revenue hit a record $11.5B (+50%) with data center doubling to $6.7B
-
Tech · 1 min readSpaceX's first report card beats across all three businesses
First earnings as a public company: revenue $7.8B vs $6.8B consensus
-
Tech · 1 min readWhat 'frontier model' means — a line that moves every year
'Frontier model' means the moving boundary of best AI capability, not a fixed spec
-
Tech · 1 min readThe EU AI Act's transparency rules just went live
Transparency rules took effect August 2 — AI-generated content must be disclosed
-
Tech · 1 min readA frontier model every two weeks: July's AI race
Google, Anthropic and DeepSeek all shipped new models within ten late-July days
-
Tech · 2 min readSK hynix lifts capex near 60% for the AI build-out
SK hynix is expanding this year's capex to the high 40 trillion won range, up about 60%
-
Tech · 2 min readMemory capex: how money becomes supply
Capex flows through construction, tool installation and yield ramp before becoming supply
-
Tech · 3 min readHBM4 vs HBM3E: what doubles, and why it matters
HBM4 doubles I/O pins from 1,024 to 2,048 and channels from 16 to 32, taking per-stack bandwidth past 2TB/s
-
Tech · 2 min readHBM4 ramp: the race turns to supply
SK hynix started HBM4 mass-production shipments in Q2 and signed long-term deals
-
Tech · 2 min readHBM explained: why stacked memory is AI's bottleneck
HBM stacks 4 to 16 DRAM dies connected by through-silicon vias to maximise bandwidth