The AI infrastructure race just got real. While giants throw billions at compute, the smartest players are solving for developer friction because the best chip means nothing if nobody can use it.
Industry moves
Huawei's open-source gamble
Huawei just did something rare in enterprise tech: admitted its platform sucks to use. At Huawei Connect 2025, Deputy Chairman Eric Xu opened with unusual candor. "Between January and April 30, our AI R&D teams worked closely to make sure that the inference capabilities of our Ascend 910B and 910C chips can keep up with customer needs."
Translation: DeepSeek-R1 dropped in January, exposed performance gaps, and Huawei spent four months firefighting. The fix is a full open-source by December 31, 2025. What's opening:
- CANN (Compute Architecture for Neural Networks): Open interfaces for compiler and virtual instruction set. Not full open-source, Huawei keeps some proprietary pieces but enough visibility for developers to optimize latency-sensitive applications.
- Mind series: Complete open-source of SDKs, libraries, debugging tools, profilers. Everything developers touch daily becomes inspect-able and community-extensible.
- OpenPangu foundation models: Full open-source, competing with Meta's Llama and Mistral AI. No specs yet on parameter counts, training data, or commercial licensing terms.
Why it matters: Huawei's acknowledging what most vendors won't: hardware performance is table stakes. Developer experience is the bottleneck. Open-sourcing the toolchain invites the community to fix what Huawei couldn't.
"Developer tools are products. If your compiler documentation requires a PhD to parse, you've already lost. Huawei's betting that transparency will build trust faster than polished marketing and they're probably right."
{{Oleksandr Koshytskyi}}
South Korea bets $390M on Sovereign AI
Seoul just launched its most aggressive AI play: ₩530 billion ($390M) split across five local companies building foundational models. The goal is to cut dependence on OpenAI, Google, and foreign tech entirely.
The Ministry of Science and ICT picked LG AI Research, SK Telecom, Naver Cloud, NC AI, and startup Upstage. Every six months, underperformers get cut. By the end, only two remain standing.
Why sovereign AI matters: It's data control, cultural context, and national security. South Korean language and culture don't map cleanly onto English-trained models. Local players argue they can build better, more relevant systems.
The challenge: Compute, talent, and capital. OpenAI burned billions getting to GPT-4. South Korea's $390M won't match that scale but it doesn't need to. Focused, domain-specific models trained on Korean data could outperform generic global systems for local use cases.
"Sovereign AI is the same logic as the sovereign cloud who controls the infrastructure, controls the data. South Korea's betting that localized models will win on relevance, even if they lose on raw parameter count."
{{Kyrylo Lazariev}}
Alibaba drops RMB 380 billion on AI infrastructure
Alibaba Cloud announced Qwen3-Max at Apsara Conference 2025: over one trillion parameters, plus new multimodal systems (Qwen3-VL for vision-language, Qwen3-Omni for text/image/audio/video). They're backing it with RMB 380 billion ($52B) over three years. The numbers:
- 600M+ downloads of Qwen series since 2023.
- 170,000+ derivative models created.
- New Model Studio platform with high-code and low-code tools for building AI agents.
What's notable: Alibaba's treating models like operating systems. CEO Eddie Wu's positioning Qwen as infrastructure, not product.
Meta poaches OpenAI scientist Yang Song
Yang Song, who led OpenAI's strategic explorations team, just joined Meta as "research principal" of Meta Superintelligence Labs.
It's the latest in a talent war that's reshaping AI leadership. OpenAI loses a strategic thinker. Meta gains someone who knows exactly what OpenAI's building next.
EU member states push for Chips Act 2.0
All 27 EU member states joined the Dutch-led Semicon Coalition, pushing the European Commission to revamp the EU Chips Act. The coalition argues the current 20% market-share goal is too vague. They want faster approvals, deeper skills investment, and a focus on securing critical technologies.
Dutch Economic Affairs Minister Vincent Karremans: "Europe's industrial strategy should adapt to the increasing geopolitical tensions in the world."
The subtext: Europe watched TSMC, Samsung, and Intel dominate while its own semiconductor industry atrophied. The Chips Act was supposed to fix that. It didn't. Chips Act 2.0 is the acknowledgment that throwing money at the problem without structural reform doesn't work.
YouTube music tests AI hosts
YouTube Music is testing AI hosts that share trivia, fan stories, and commentary about tracks. Think Spotify's AI DJ, launched two years ago, but for YouTube's ecosystem.
Why it matters: Personalization at scale. Human DJs don't scale to 600M users. AI does. The question is whether they feel authentic or algorithmic. Spotify's AI DJ walks that line. YouTube's trying the same play.
Product reality check
OpenAI's first big ChatGPT campaign
OpenAI just dropped its first major brand campaign: 35mm film, real people using ChatGPT to cook, learn, travel, and hit fitness goals. No dystopian robots. No sci-fi panic. Just "everyday magic," per Elke Karskens, OpenAI's UK head of marketing.
"This is smart positioning. OpenAI's battling perception as much as competition. The scariest thing for mass adoption is that people think it's too complicated or creepy. Shoot it on film, show real use cases, strip the hype. It's brand storytelling 101, executed perfectly."
{{Oleksandr Koshytskyi}}
Bedtime stories app: accessibility meets AI
The Royal Society for Blind Children partnered with Innocean Berlin to launch Bedtime Donations, an app that turns parents' voices into audiobooks for visually impaired children worldwide.
How it works: Parents read from a curated list of stories. The app records their voice, runs quality checks, and uses AI to turn recordings into free audiobooks.
"This is what good product design looks like, identify the constraint (limited audiobooks for blind kids), remove it (crowd-source voices + AI processing), and make it dead simple to use."
{{Danylo Dubrovsky}}
Design wins
Lazarev.agency delivers for 11Sight conversions through storytelling
11Sight needed better conversions and click-through rates. They came to Lazarev.agency for a full website redesign, new design system, and digital experience overhaul.
Our approach: We fused storytelling and animation to reflect the product's innovation while keeping it human. To capture sales, we offered a free trial to connect with 11Sight reps via their own video call platform. Let the product sell itself.
The principle: Show, don't tell. If your product is a video call platform, use it in the conversion flow. If your product is innovative, the site should feel innovative. Match form to function.
Lazarev.agency client success updates
Redbrain gets a E2E acknowledgement and moves to Northspring
Jared Owen, CEO of Google shopping business Redbrain, reflected on being included in the E2E Tech 100. He called it "a great benchmark for the business to understand how it's performing in the tech space."
Moreover, Redbrain just moved into a bigger office at Northspring in Birmingham to support growth and collaboration. Owen: "On walking into the reception area, you could feel that the Northspring building was of a very high quality for the area."
Dynascale is pushing hard
Dynascale, a Lazarev.agency client that we redesigned, is already pushing hard. Their latest results:
- 2x performance increase.
- 100% platform availability.
- 2x cost reduction.
Dynascale's turnkey migration helped CashCall Mortgage focus on what matters: delivering service, not managing infrastructure.
"Good design is about clarity. Dynascale's users needed to understand complex migration processes without getting lost. We stripped the noise, rebuilt the UX around decision-making, and the results speak for themselves."
{{Kyrylo Lazariev}}
Events to monitor
- GAI World 2025 (September 29-30, Boston, MA) – Applied AI in business with talks, panels, and case studies. Best for executives, product leads, and strategy teams.
- MLcon New York (September 29-October 3, New York, NY & Hybrid) – Practical ML and MLOps. Learn to build, deploy, and scale ML systems with real-world tools.
- 7th Annual Machine Learning in Quantitative Finance (September 29-October 1, New York, NY) – For financial institutions focused on trading, modeling, and risk management.
- SDV & AI in Automotive USA 2025 (September 30-October 2, Ann Arbor, MI) – Focused on software-defined vehicles and AI in the automotive industry.
This week's reality check
How innovative e-commerce sites use design to drive sales
Nike doesn't dominate e-commerce because of the swoosh. It dominates because the digital experience feels effortless and intentional. Personalization, frictionless mobile UX, and design that converts. The fundamentals:
- E-commerce is the new retail standard: Online retail sales will surpass $4.3 trillion by 2025.
- Personalization boosts performance: AI-powered recommendations and dynamic previews turn high-intent visitors into loyal buyers.
- Frictionless mobile UX wins: 80% of retail traffic comes from mobile. Responsive design is the baseline.
"E-commerce design is conversion optimization disguised as aesthetics. Every pixel either moves users toward purchase or friction. Nike nails it because they obsess over micro-interactions product zoom, size selection, checkout flow."
{{Oleksandr Koshytskyi}}
What's сoming next week
More AI infrastructure plays, more regulatory chess moves, and hopefully fewer vague "AI-powered" product launches that don't actually do anything.
🔥 Stay sharp. Stay with Lazarev.agency.
