Tech news picked on 2026-06-25
Europe Puts the Brakes on Washington's Chip War, Even Allies Resist Control
- The Netherlands openly opposed the US MATCH Act targeting ASML, with its trade minister personally visiting Washington.
- ASML earns about one-fifth of its system sales from China.
- As the period when allies lined up uniformly under US control comes to an end, regulatory scope is now tied to alliance negotiations and exception design.
- Teams across the semiconductor supply chain should prepare for the moment when clients move before the law is finalized.
From Scanner to Agent, Daybreak Breaks the Moat of Security Startups
- OpenAI has expanded its cybersecurity program Daybreak by introducing the Codex security plugin and GPT-5.5-Cyber.
- GPT-5.5-Cyber directly generated 8 proof-of-concept (PoC) information leak exploits and 24 local privilege escalation exploits in the Linux kernel.
- This marks a transition from being a scanner that just raises alerts to an agent that finds, verifies, and patches vulnerabilities.
- For security startups, it is now the time to critically ask whether their product is genuine, focusing not just on detection screens but also on the context of the customer environment and human review.
Qualcomm Bought Not Chips but a Compiler for $3.9 Billion; AI Moats Shift from Silicon to Software
- Qualcomm is acquiring Modular for $3.9 billion entirely in stock, which allows running models on NVIDIA, AMD, Intel, and custom ASICs without rewriting code.
- Modular is the company behind the Mojo language and MAX inference stack.
- A chip company purchasing a hardware-independent compiler signals that moats are moving from silicon to the software layer above.
- To save inference costs, organize your system so you can switch backends when needed.
As Agents Begin to Work, the Real Market Opens with 'Control'
- On June 23, Zafin unveiled AIOS, an AI agent control tower designed for regulated financial institutions.
- This product integrates agent registration, permission restrictions, human decision points, cost control, and evidence recording into a single workflow.
- As with DevOps and MLOps, as automation increases, competition shifts from features to operational accountability.
- If you plan to sell agents in regulated industries, design control not as a cost added later but as a feature that sells from the start.
Employee surveillance data leaked to employees, backlash against monitoring SaaS
- Meta has paused its MCI program that collected employees' keyboard inputs, mouse movements, and screens for AI training.
- This was due to an incident where collected private conversations, personnel evaluations, tax, and medical records were exposed to all employees within the company.
- Products that require accumulating a large amount of sensitive original data to gain value can have their entire trust destroyed by a single small permission mistake.
- For employee monitoring SaaS, it's necessary to ask first whether value holds even if less data is collected rather than relying solely on security promises.
Bank Lines Up AI and Deep Tech Entrepreneurs, Signals from Dinolab
- Woori Financial Group discovers not only fintech but also AI and deep tech non-financial companies through Dinolab.
- Over nine years, they have nurtured 199 companies with direct and indirect investments totaling 371 billion KRW.
- Introducing to affiliates is a strong business asset, but if exclusivity clauses and IP ownership are unclear, the product can become confined to a single client's operations.
- It is better to write down key reusable features and non-negotiable boundaries in sentences before submitting applications.
Qualcomm Sells Server CPUs to Meta, Founders' Inference Cost Shaken Again
- Qualcomm unveiled its own data center CPU, Dragonfly C1000, and secured Meta as its first customer.
- Mass production is in 2028, with double the performance per watt and inference accelerators up to 8 times that of GPUs, according to the company.
- While it's welcome that cracks appear in the Nvidia and x86 duopoly, there's no guarantee that API prices will drop just because chip efficiency improves.
- For small teams, having the option to switch suppliers regardless of who leads is an asset.
NVIDIA's 45°C Hot Water Cooling Cuts AI Data Center Water Consumption to Nearly Zero
- NVIDIA has unveiled a 100% liquid cooling design running at 45°C hot water.
- It circulates the cooling water filled once over the facility's lifespan, reducing water consumption, previously about 2.6 million gallons per MW per year, to nearly zero, and the 50MW pilot plant in Finland uses waste heat to warm about 20,000 homes in Espoo.
- The opportunity for entrepreneurs lies more in businesses connected outside the data center such as cooling water circulation, water treatment, and waste heat sales, rather than the cooling equipment itself.
GPUs have become geopolitical assets, with compute access a structural constraint for startups
- According to the FT, Nvidia AI chips blocked by the US are sold on China's black market at roughly double the price.
- The DGX B300 costs about 8 million yuan, more than twice the US price, and $1 billion worth has been smuggled in over three months following tightened export controls.
- While controls do not completely block GPU movement, they impose higher costs and unstable supply on those seeking compute legally.
- Designs that avoid dependence on a specific chip and single supply source determine survival.
Frontier Models Leak Through Distillation, The Real Moat of Companies Built on APIs
- Anthropic has accused Alibaba.
- They claim that from April 22 to June 5, they used 20,0005 thousand fake accounts to interact with Claude 28.8 million times, extracting the model's capabilities.
- If frontier models can be matched through distillation just by collecting enough outputs, then simply having the latest model is not a moat.
- Let's consider whether our product has elements that cannot be taken just by gathering outputs, such as proprietary data, positioning in workflows, and distribution.
Memory supercycle proven by Micron's income statement: now chip prices set AI unit costs
- Micron has forecast a record high for 2026 fiscal Q3 revenue of about $33.5 billion and a non-GAAP gross margin of 81%.
- HBM grew from less than 5% of DRAM revenue in 2022 to more than 30% in 2026, and only three companies mass-produce it.
- If you built pricing and margins on the assumption that model price lists stay the same, it is time to review them.
- Set a ceiling on the inference cost one AI feature can bear, and write down what to fix first if it goes over.
Memory Locked in 3–5 Year Cycles, Hardware Founder's BOM after Annual Negotiations End
- Micron secured a minimum revenue floor of $100 billion with only 14 out of 16 strategic customer contracts, and 20% of DRAM supply is tied up in long-term contracts.
- Samsung and SK Hynix also switched from annual contracts to 3–5 year agreements.
- When large customers pay advances to lock in supply first, those costs pass through component prices and cloud inference pricing to smaller companies.
- For hardware teams, memory has become more a matter of financing than purchasing.
Frontier Image Model Released, Moat Shifts from Model to Workflow
- Krea AI released the 12 billion parameter image model Krea 2 as open weights on June 22.
- The Turbo version can generate 2K images in about 2 seconds on consumer hardware.
- A bigger change than rankings or speed is that you can run and adjust the weights yourself.
- As the gap in generation quality narrows, the competition shifts to how deeply the solution integrates into customer workflows and how much unique data and distribution have been accumulated.
The 30% commission is broken: How opening external payments on Google Play changes the app economy
- Google will open external and alternative payments on the Play Store in the US, UK, and EEA starting June 30.
- As a result of the Epic lawsuit settlement, the standard commission will decrease from 30% to 20%, with some rates lowering to 9–10%.
- However, immediately switching to external payments is not the right sequence.
- Handling refunds, taxes, and payment failures directly may cause smaller apps to incur higher operating costs than the fees saved.
- Only consider changing payment flows after evaluating actual transaction size and operational capacity.
When Silicon Valley Cheers for Chinese Models, How OpenWeight Changes Startup Cost Sheets
- DeepSeek V4 charges $3.48 per one million output tokens, while for the same task OpenAI charges $30 and Anthropic charges $25.
- This is why Silicon Valley is excited about Chinese models that allow running weighted operations directly.
- However, switching every time a cheaper model appears is not the answer.
- Since direct serving operational costs and regulatory reviews remain, it's better to secure negotiation power by structuring your system to be able to switch between any model first.
Inference chip margins took a hit, and Cerebras's 47%→38% plunge is a warning
- Cerebras shares fell nearly 20% after the earnings release.
- Its gross margin of 47% in Q1 led it to cut full-year guidance to 38~41%.
- A shortage of data center space forces it to rent back its own systems and hand them to customers, which cuts margins by 10~15% points.
- Even as chips improve, physical bottlenecks such as space remain.
- Leave out of business plans the assumption that inference prices will only fall in one direction.
Regulated Bank Equipped 120,000 Employees with AI, The Real Gateway to B2B AI
- Spanish bank BBVA is expanding ChatGPT Enterprise deployment to 25 countries covering 120,000 employees.
- This decision follows a pilot involving 3,300 employees during May 2024, growing to 1,000 10,000 employees.
- The conservative regulated bank has effectively shifted AI from a pilot initiative to a core operation.
- However, Korean financial companies must address data storage, authorization segregation, and audit compliance before contracts are signed, so teams familiar with the approval processes of a given task will onboard faster in the domestic market.
MacBook and iPad Price Rise by 20%, Memory Supercycle Hits Startup Costs
- Apple raised MacBook and iPad prices by up to 20% on June 25.
- AI data centers consuming memory caused DRAM and NAND prices to rise sixfold over the past year.
- Large companies can secure bulk orders and adjust replacement timing, but small teams risk destabilizing cash plans even if the purchase timing of development equipment and servers shifts once.
- Decide what to buy now and what to lease based on utilization rates.
Hybrid SaaS Proven by Jamaeek: Reattaching Human Touch to AI Translation
- Voice루's video translation platform Jamaeek launched a premium plan bundling six features involving human work, such as dedicated reviewers and automatic application of customer-specific guides.
- It's a hybrid model that adds expert quality on top of AI speed.
- As AI gets faster, humans disappear but move to more expensive stages like review and guarantee.
- However, this structure only works in areas where the failure cost is high enough customers pay extra for the guarantee.
Buy Guardrails When Installing Agents, A New B2B Market for Safety Tools
- NVIDIA released the content safety model Nemotron 3.5, and RIFT-Bench published results of automated attacks on 45 agent systems.
- As companies deploy agents, runtime safety classifiers, guardrail APIs, and red teaming are solidifying as product categories.
- However, selling generic classifiers or one-off red team reports as-is is problematic.
- Customers pay based on how accurately they enforce their own policies and how long they stay engaged with the operating agents.
A New Challenge in IT Services: Vishal Sikka's Startup
- Former Infosys CEO Vishal Sikka has started an IT services startup with professionals from SAP and Infosys.
- Large IT services firms can quickly integrate new technologies into existing contracts thanks to their workforce and customer networks, so when a small team competes in the same scope, only the difference in scale is apparent.
- More than assembling experienced professionals and capital, the key to winning will be how clearly they solve the narrow and complex problems that large companies have overlooked.
Unicorn Bridge Selects Teams to Go Global
- Sage and LifeX aim at overseas markets using AI specialized in specific industries, leveraging support programs like Unicorn Bridge.
- It's a mistake to confuse being selected for a support project itself as a global expansion strategy.
- First, narrow down to a single target industry and its field problems, then organize AI talent and cooperative networks suitable for that.
- Concrete results in one industry persuade both evaluation and expansion more than just saying the market is large.
Startup Trends Driven by AI Manufacturing Innovation and Local Investment
- At the 2nd Sejong Startup Venture Forum, discussions centered on AI manufacturing innovation and the local investment ecosystem, along with the Madup IPO and advice from a YC-affiliated founder.
- As conversations about technology and investment grow, entrepreneurs must return to the voice of the customer.
- This week, choose one customer segment to validate, and verify the realness of the problem not by surveys but through direct conversations.
- Local funding and AI serve as tools to aid judgment, not to decide the product for you.