Daily Newsletter Digest

Cerebras IPO Repriced to $150-$160 as Demand Surges 20x Over Supply

May 11th 2026

Today's News

Cerebras IPO price range raised to $150–$160 per share, targeting $4.8B as demand surges

Techmeme

Cerebras Systems is raising its IPO price range to $150–$160 per share (from $115–$125) and increasing share count to 30 million, potentially raising ~$4.8B at the top of the range ahead of May 13 pricing. The company has drawn orders for more than 20 times the available shares, reflecting strong investor appetite for AI infrastructure hardware. Cerebras' chips are optimized for inference workloads, positioning the company to benefit from the industry's shift from model training to deployment. The company has secured Amazon and OpenAI as customers since its previous IPO attempt in 2024, which was withdrawn due to CFIUS review of its partnership with UAE-based G42.

ByteDance boosts 2026 capex to $30B+, up 25%, as AI infrastructure costs and memory chip prices surge

Techmeme

ByteDance is increasing its 2026 capital expenditure to over 200 billion yuan ($30B), a 25%+ jump from preliminary plans, driven by rising AI infrastructure commitments and escalating memory chip costs. The company is allocating a proportionally larger budget to domestic AI chips to mitigate geopolitical risk and comply with Beijing's push for domestic semiconductor adoption. ByteDance is also expanding aggressively overseas, including a $25B investment in Thailand's data infrastructure and an additional €1B for a Finnish data center, underscoring the capital intensity of competing in frontier AI model development.

TSMC delays High-NA EUV adoption until 2029, signaling potential reversal of semiconductor cost-down economics

Azeem Azhar, Exponential View

TSMC has announced no plans to deploy ASML's High-NA EUV lithography equipment through 2029, citing prohibitive costs—the machines cost upwards of €350M ($410M) each. Deputy Co-Chief Operating Officer Kevin Zhang stated that TSMC's leading-edge A13 chip will not enter production until 2029. This decision represents a critical inflection point: while transistor density per wafer-dollar improved reliably for a decade following EUV adoption in 2019, High-NA economics have failed to restore the historical cost-down curve that enabled five decades of 18–24 month improvements. If semiconductor manufacturing can no longer deliver predictable cost reductions, it fundamentally constrains the infrastructure economics underlying AI model scaling and deployment efficiency gains—the foundational assumption that has powered the cloud computing and AI era.

Google's Decoupled DiLoCo enables resilient distributed training across global datacenters at scale

Jack Clark from Import AI

Google demonstrated training a 12B parameter model across four U.S. regions using standard inter-datacenter connectivity (2–5 Gbps), maintaining 88% training efficiency even under aggressive hardware failures versus 58% for traditional methods. This capability signals how compute superpowers can consolidate global infrastructure into unified training systems, creating competitive advantages in large-scale model development.

Worth Reading

Cerebras Systems Raises IPO Price Range to $150-$160 Per Share Amid Surging AI Chip Demand

Reuters

I read the full Reuters report on Cerebras' IPO repricing. Main takeaways: - IPO price range raised to $150-$160/share (from $115-$125), with 30M shares (vs. 28M), targeting ~$4.8B raise—largest global IPO YTD - Demand exceeds supply by 20x+, driven by AI inference shift: Cerebras chips better suited for inference than GPUs for model deployment - Customer roster now includes Amazon and OpenAI alongside G42 (which previously represented 80%+ revenue); CFIUS cleared the G42 deal - Second IPO attempt after 2024 filing pulled due to national security review; company now positioned as inference-focused alternative to Nvidia in training-saturated market

Read more

TSMC Delays High-NA EUV Adoption Through 2029 Due to €350M Equipment Cost

Bloomberg

I read the full report. Main takeaways: - TSMC, ASML's largest customer, will not deploy high-NA EUV lithography machines through 2029 due to cost (€350M+ per unit), signaling price resistance at the leading edge of chipmaking equipment - TSMC's A13 chip production delayed to 2029, pushing next-generation node deployment timeline further out than previously expected - Equipment cost economics are now constraining adoption velocity even for the world's most advanced chipmaker, with implications for AI chip production timelines and capex efficiency across the industry

Read more

Neural Computer: A New Machine Form Is Emerging

MetAuto

I read the full essay. Main takeaways: - Proposes a new computational paradigm (Neural Computer) where models take on runtime responsibilities currently handled by program stacks and control layers, rather than just serving as agents using conventional computers. - Mature Neural Computers may diverge from dense foundation-model scaling (1B-10T range) toward sparser, more addressable 10T-1000T substrates with discrete, routable components inspired by NAND-style logic rather than biological brains. - Capability upgrades shift from wholesale parameter replacement to incremental installation of reusable neural routines into persistent internal state, enabling composition and invocation patterns closer to memory and callable subprograms than traditional model swaps. - Three concurrent trends enable this: agents achieving production reliability (Cursor, Claude Code as defaults), world models entering real closed loops (Waymo simulation, NVIDIA Cosmos), and conventional computers showing structural friction under open-ended, long-horizon AI workloads. - Runtime becomes the organizing principle—what persists, what changes the machine versus rewriting it, and how capabilities remain installable, reusable, and governable within the learned system itself.

Read more

Articles Read