Tech news picked on 2026-06-24
- Menlo Ventures raised its largest fund ever in 50, amounting to $3 billion.
- The bet of investing over $500 million into Anthropic in 2024 returned about $14 billion equity.
- As capital flows more toward foundation models and adjacent infrastructure, founders outside that space may have follow-on funding halted regardless of business results.
- Companies with their own data, distribution, and low capital consumption keep options, rather than forcibly riding model competition.
OpenAI Acquires Gitpod-Originated Ona, Controlling the 'Execution Environment' of Agents
- On June 11, OpenAI announced the acquisition of cloud development environment startup Ona, formerly Gitpod.
- They explained that as weekly Codex users surged from 3 million in April to 5 million in June, the execution environment to run models became a bottleneck.
- The thin wrapper functions get absorbed into the platform faster than founders can fix their products.
- The key battleground is the regulatory and operational gaps the platform fills late , such as permissions, audits, and data boundaries.
OpenAI Makes Its Own Chip, A Bottom Line on AI Cost for Founders to Rethink
- OpenAI has unveiled its first in-house inference chip, Jalapeno, which the company says outperforms existing GPUs in watt-per-performance and costs about half as much.
- However, even if chip costs decrease, there is no guarantee the savings will go to domestic startups.
- Since overseas model operators control pricing and access, the negotiation power gap may widen first.
- Korean teams must rely on designs that allow model changes and assets built on domestic language, industries, and regulations.
Bottleneck moved from GPU to memory, meaning SK Hynix sold all 2026 supply
- SK Hynix has virtually pre-sold all 2026 production volume of DRAM, HBM, and NAND.
- Quarterly operating profit is about 11.4 trillion won, a record high, up about 62% from the previous year.
- Being close to the supplier does not mean domestic startups get supply first.
- In a structure where large customers bundle production, domestic teams are also secondary demanders.
- The advantage for Korean teams lies not in cheaper memory, but in designing product structures that use less memory first.
Open model matches GPT-5.5 at 1/6 cost, moat lies outside the model
- On June 17, Z.ai released GLM-5.2, which outperformed GPT-5.5 on several long-horizon coding benchmarks at about one-sixth the cost.
- It features 753B MoE, 1M context length, and has its weights released under the MIT license.
- Companies relying on closed APIs need to reconsider build versus buy calculations.
- Select a core task to reevaluate quality, cost, and data export terms directly, and verify whether switching models retains product value.
- Tesco is removing 40,000 server workloads from VMware.
- They bought perpetual licenses in 2021, but after Broadcom's acquisition, faced a 175% price hike and a demand for 23.5M dollars over 1 years, leading to a 100 million pound lawsuit.
- This exposes a pattern of forcing already paid software into subscriptions again.
- In migration businesses, it is important to catch the moment the customer decides to leave rather than the day the migration ends.
Samsung Hands ChatGPT to All Employees, The Gaps Created by In-House AI
- Samsung Electronics rolled out ChatGPT Enterprise and Codex to all domestic employees and global staff in the DX division on June 21.
- The company completely reversed its stance, having blocked generative AI three years ago due to a source code leak.
- However, installing general-purpose AI enterprise-wide does not automatically lead to work innovation.
- Success depends on how tightly permissions, security, internal data, and workflows are connected, leaving room for deep, focused products to fill the gaps.
AI Agents Have Entered the Workplace: Three Signals That Broke Out on the Same Day
- On June 23, Menlo Ventures' $3 billion fund, Claude Tag residing inside Slack, and MoEngage's marketing agent bet were announced together.
- This is a signal that enterprise agents are moving past demos into actual work.
- The bottleneck is closer to the ability to explain authority and responsibility than model performance.
- When selling agents, entering a narrow product that shows execution scope and audit logs together reduces obstacles in the diffusion phase.
EA's $55.5 Billion Private Acquisition: How Saudi Capital is Buying into Gaming
- EA has entered its third round of layoffs this year ahead of a $55.5 billion acquisition led by Saudi sovereign wealth fund PIF.
- This round targets recruiting, customer support, trust and safety, and IT departments.
- This indicates that cost-cutting efforts are underway before the deal closes.
- While it will take longer to reveal what strategies the new capital will push, smaller game teams would benefit from focusing on available talents and gaps in the market during this time.
ChatGPT Launches Ads, Consumer AI's Revenue Model Ultimately Settles on Advertising
- OpenAI has set up its first booth at Cannes Lions and officially started selling ChatGPT ads to marketers.
- This signals that consumer AI, which has been bearing the burden of free users and inference costs, ultimately converges on advertising.
- For services discovered within AI assistants, free reach is more of a temporary privilege than a right.
- Set the boundary between recommendations and ads as a product principle, and cultivate at least one channel where users come directly.
Bank of England Lifts Holding Limits and Imposes Issuance Ceilings: New Rules for Pound Stablecoins
- On June 22, the Bank of England released a draft of the pound stablecoin rules.
- The individual holding cap of £20,000 and corporate cap of £10 million were removed, but a £40 billion issuance ceiling per product was set, with up to 70% of reserves allowed to be invested in government bonds to earn interest.
- Rather than deregulation, the control target shifted from users to issuers.
- Payment startups need to first decide which currency and jurisdiction to choose before focusing on technology.
$500 Billion Solidifies into Concrete: What Compute Buildout Leaves for Founders
- OpenAI, SoftBank, and Oracle's $500 billion Stargate project is solidifying into a physical presence.
- The Texas Abilene hub is already operational and approaching 7 gigawatts with five new sites.
- Even as supply grows, there is no guarantee all founders get compute at the same prices and terms.
- Don't leave inefficient structures unchecked expecting compute to get cheaper; establish flexible systems first.
Token Prices Fell 280x but Bills Rose 320%: How to Understand AI Cost Polarization
- DeepSeek cut the V4-Pro token price by 75%, while Meta raised AI infrastructure investment to $145 billion by 2026.
- Token prices have dropped 280 times in 2 years, but enterprise AI bills have increased by 320%.
- This means that as prices drop, usage rises faster, leading to larger bills.
- Instead of immediately seeing the price cut as margin improvement, focus first on redesigning to reduce calls and improving caching.
The Korean Startup Ecosystem from a Founder's Perspective
- This article discusses how the Korean startup ecosystem has gained new opportunities thanks to the Ministry of SMEs and Startups' deregulation of venture investment and the reorganization of funding support systems.
- Interest has increased in personalized health solutions and cloud services.
- However, even though the funding gate has widened, investors do not create the product's persuasiveness for you.
- You must decide what to prove first among product-market fit, measurable customer outcomes, and validated cases to see the next funding round.
Integration of AI and Art: Opportunities and Challenges for Entrepreneurs
- Korean Deep Learning has emerged prominently, surpassing Google and OpenAI, while the Chaos Foundation addresses the future of AI through concerts that combine art and science.
- Money tends to flow first to clear problems with definite buyers, such as corporate document processing, rather than slogans about the fusion of AI and art.
- Grants can buy development time but do not guarantee customers who will generate revenue.
- Even fusion experiments become viable businesses only when there are answers to who pays and why.
AI and Cloud Innovation: An Entrepreneur's Perspective
- Data Alliance, which operates the distributed GPU cloud G-Cube, has been selected as a leading-edge startup in the AI sector.
- The selection and support do not mean business viability has been proven, but rather that access to funding and networks has expanded.
- The article lacks details on cost savings or customer outcomes.
- For businesses like GPU cloud subject to price comparison, it is better to first create numbers that show competitive advantages rather than just the AI label.