DanielRomero

People familiar with the ASIC industry said $GOOG multi-vendor strategy is aimed at preventing any single partner from gaining too much pricing power, particularly $AVGO
For Broadcom, the main blow is losing its status as Google’s sole long-term partner
Broadcom primarily handles flagship TPUs used for AI training, with its agreement running through 2031. While $MRVL deal does not appear to affect training chips, its NICs and storage controllers will compete directly with Broadcom’s networking products

Valuation per MW of Data Center Operators
$DGXX is the cheapest, $WYFI is the most expensive, and $IREN is the cheapest neocloud
• Digi Power X | $DGXX: $0.2B EV / 400 MW = $0.5M EV/MW
• Soluna | $SLNH: $0.2B EV / 206 MW = $1.1M EV/MW
• Bitdeer | $BTDR: $4.4B EV / 2,825 MW = $1.5M EV/MW
• HIVE Digital Technologies | $HIVE: $0.9B EV / 564 MW = $1.6M EV/MW
• MARA Holdings | $MARA: $3.6B EV / 1,900 MW = $1.9M EV/MW
• CleanSpark | $CLSK: $3.7B EV / 1,800 MW = $2.1M EV/MW
• IREN | $IREN: $16.8B EV / 5,800 MW = $2.9M EV/MW
• KEEL Infrastructure | $KEEL: $2.2B EV / 648 MW = $3.4M EV/MW
• Riot Platfor

$META Is Now One of $MSFT Largest AI Customers
Meta is spending hundreds of millions of dollars per year on AI models through Azure Foundry
According to Bloomberg, Meta consumes trillions of tokens every week, using external models including OpenAI’s to support software development and even evaluate outputs from its own models
Microsoft’s AI revenue is still heavily concentrated among other technology companies. ByteDance is reportedly Foundry’s largest customer, while Adobe, Perplexity and Sierra are also major users. OpenAI alone accounted for around 70% of Microsoft’s AI revenue in its late

$GOOG Is Moving to a Multi-Vendor Custom Silicon Strategy
Current $GOOG silicon split:
• Sunfish | TPU v8, training-focused | $AVGO
• Zebrafish | TPU, lower-cost inference-focused | MediaTek
• Merope | LPU / inference accelerator | $MRVL
As DIGITIMES analyst Luke Lin said, “Marvell is not taking over the existing TPU product lines. The scope described in the filing sits around the TPU ecosystem rather than inside it.”
Marvell also won programs for:
• Storage controllers
• NICs
• Memory interface controllers
• Near-memory compute
DIGITIMES says Marvell’s accelerator work could include Merope,

$SKHY is moving closer to $NVDA and $AMD
The company is building an HBM design team in San Jose to co-design next-gen HBM directly with major US AI chip customers
San Jose is next to Santa Clara, where Nvidia and AMD are headquartered, while Broadcom also has major operations in the area
Having HBM design staff nearby allows faster coordination during next-generation product development and quicker feedback to SK Hynix’s R&D teams in South Korea

$NBIS announces pricing of upsized private offering of $5.0B of convertible senior notes
• $3.0B of 0.50% notes due 2030
• $2.0B of 4.50% notes due 2034
Conversion prices of $313.46 and $324.65, representing 40% and 45% premiums
Effective conversion prices at maturity of $344.81 and $405.82, representing 54% and 81.3% premiums
• Option for an additional $750M
Net proceeds are expected to be $4.94B, or $5.68B if the additional notes are fully purchased
Nebius will also exchange $800M of existing convertible notes for ~15.8M shares

Panel-level advanced packaging is exploding
According to Yole’s 2026 report, the advanced packaging market is expected to post a 9.1% CAGR from 2025 to 2031
PLP is growing much faster
Yole expects PLP revenue to rise from $312 million in 2025 to $2.959 billion in 2031, representing a 45.5% CAGR over the same period

$NBIS to Raise $4.5B Through Convertible Notes
Prepayments and GPU-backed loans are important, but such an aggressive buildout requires more than that
Predictable and understandable. The market cap is very high at the moment, so it’s the perfect time for this
Crazy that they’re raising the equivalent of their entire market cap when they went public

DigiTimes reports that $AMD is likely cooperating with $GOOG on CPU IP, programmable logic, interconnect technology, and/or advanced packaging capabilities
The reason is AMD’s experience integrating CPUs and accelerators in the same package
The disruption to $AVGO is limited

$INTC Advanced Packaging Plant in Malaysia Now Online, According to SemiAnalysis
A growing share of South Korean HBM shipments is now heading to Malaysia, while shipments to Taiwan have fallen sharply
Around $1.3 billion worth of HBM is now being exported to Malaysia, compared with less than $3 billion to Taiwan, down from more than $5 billion just a few months ago
Since TSMC has no advanced packaging facility in Malaysia and other local players remain relatively small, SemiAnalysis argues that Intel is the most likely destination
Intel recently completed the final phase of its $7 billion Proj

Anthropic ARR:
> December: $9B
> May: $47B
> July: $65B
According to Reuters
Anthropic is projecting $190B to $200B in revenue by 2028

$NBIS Vineland Phase 2 Data Center is now Approved
Phase 1 of the data center is already under construction, and Phase 2 will add 600,000-square-feet to the project, that is hoping to ultimately take up 2.6 million square feet
Board members who voted to approve the second phase said that the company behind the data center, Data One, has been responsive to resident complaints and they will continue to watch over the project
Data One has previously told NBC10 that it supports city leaders and appreciates the patience of the Vineland community. The company also said the site is operating in compl

Bottlenecks keep constraining the AI buildout
According to ETNews, some Samsung Electro-Mechanics high-capacitance MLCCs are seeing average lead times of around 40 weeks, while Murata products of 1 microfarad or more reached 30 weeks in July, up from 24 weeks in June, with some distribution channels quoting up to 36 weeks
AI servers can require five to 13 times as many high-capacitance, compact, and high-reliability MLCCs as conventional servers. Some capacity expansions previously expected in Q4 2026 have also been pushed into 2027, limiting near-term supply relief
The extended lead times are

Samsung’s China Sales Triple, Surpassing US Sales
Samsung’s sales to China surged 208% YoY to KRW88.6T in H1 2026, overtaking the US at KRW70.7T, as the AI-driven memory upcycle boosted demand and pricing
The semiconductor DS division generated KRW209.2T in revenue, 68.5% of Samsung’s total, and 97.4% of operating profit. Samsung’s DRAM market share also rose to 39.4% from 34.0% in 2025, helped by AI server demand and higher memory prices
Samsung has begun selling HBM4 for Nvidia’s Vera Rubin platform, but Nvidia was still not among its five largest customers in H1. The top five were Alphabet,

$NVDA Nears $100B Credit Guarantee for OpenAI’s 10GW Ohio Data Center
Nvidia is nearing a deal to provide around $100B in credit guarantees for OpenAI’s planned 10GW Ohio data center, down from earlier discussions of as much as $250B
The first phase could cost $250B over two years, while the full campus is expected to require at least $500B, including chips, power, construction and other infrastructure
The project is being developed by SB Energy, backed by SoftBank. Nvidia is also considering investing $3B in SB Energy, on top of its existing exposure to OpenAI
The guarantee would help SB Ener

$AMD Taalas Acquisition Aims to Reduce HBM Dependence
AMD is acquiring Taalas, an AI chip startup founded by former AMD and Nvidia architect Ljubisa Bajic, as it expands beyond GPUs into more specialized AI inference architectures
Taalas takes a radically different approach by hardwiring model weights directly into silicon instead of storing them in external HBM. The company says this allows its systems to operate without HBM, advanced packaging, 3D stacking or liquid cooling
Its first HC1 chip, built on TSMC 6nm, reportedly runs Llama 3.1 8B at 16,960 tokens per second per user, 48x faster th