VARA AI Fund had 9 employees and about $26.4B under management at the end of June, 16 months after launching its hedge fund. Rumor going around today that it blew up. Their 3/31 13F book at today’s prices is $13.1B across 60 names. $NBIS $1.28B $BE $1.27B $LITE $1.03B $CRWV $889M $STX $728M $VRT $720M $WDC $624M $CLS $537M $COHR $492M $FIX $426M

BOFA WARNS AI BOOM COULD FOLLOW HISTORIC BUST CYCLES Bank of America warns that the AI investment boom is showing signs of a classic capital-cycle risk: surging spending, intensifying competition and falling prices. Hyperscaler capital expenditure expectations have jumped from $330 billion to $1 trillion, equivalent to roughly 3% of U.S. GDP. Meanwhile, AI token prices have fallen around 50% since summer, raising concerns that growing usage may not translate into sufficient revenue. The key risk: AI technology could succeed while investors suffer from overinvestment and declining returns.

DAN IVES CALLS AI STOCK SELLOFF A CLEAR BUYING OPPORTUNITY Dan Ives says the sharp declines in CoreWeave ($CRWV ), IREN ($IREN ) and Nebius ($NBIS ) represent buying opportunities rather than warning signs. Despite rising borrowing costs, Ives argues AI computing demand remains strong, with chip demand reportedly exceeding available supply by 13 to 1. He expects 800–1,000 data centers to be built in the coming years. His key message: Higher interest rates make AI infrastructure more expensive to finance, but do not necessarily weaken demand.

At the same time, as bullish as I am on some of the names I talk about often, if you are low on cash-raise cash. Doesnt matter how much you like a company-survival is the name of the game. I'd rather own a smaller position for 3 years to let the story play out instead of being over leveraged and panic selling when your back is up against the wall one day.

$CBRS The Valuation Hawks Are Running Out of Excuses At roughly $35B in market cap today, I see an S-tier chipmaker with exceptional wafer-scale IP, a valuable ultrafast-inference niche and enormous demand. For my 2029–2030 thesis, this looks cheap. The commercial evidence is substantial: Q2 core revenue grew 103%, cloud revenue nearly quadrupled, and remaining performance obligations reached $25.4B. Management plans to more than triple revenue in 2027. OpenAI, AWS $AMZN, AMD $AMD and CrowdStrike $CRWD are already part of the customer and partner story. On September’s Latent Space podcast, CTO Sean Lie said Cerebras was “sold out of everything that we’re building.” He described OpenAI allocating capacity to critical research and incident response. That illustrates how valuable speed becomes when every second matters. Reuters reported that Jensen Huang’s $NVDA approached Cerebras about a potential acquisition and Cerebras declined. Asked about acquisition approaches, Feldman’s answer was classic: “We don’t kiss and tell. We’re gentlemen.” $NVDA’s roughly $20B Groq licensing-and-talent deal makes $CBRS at approximately $35B in its first year as a public company look like a serious bargain to me. For less than twice that strategic price tag, investors get exposure to exceptional wafer-scale IP, established OpenAI and AWS $AMZN relationships, and rapidly growing inference demand. I see an opportunity to own a future powerhouse early in its public-market life. Further swings lower could make that entry price even more compelling. The model-specific output-speed comparisons have real punch: -Llama 4 Scout: 2025 benchmark figures reproduced by Cerebras showed approximately 2,014 tokens/sec versus Groq’s 275, about 7.3×. -Llama 3.3 70B: Cerebras’s September 2025 comparison reported over 2,500 versus approximately 403, over 6×. -GPT-OSS-120B: today’s Artificial Analysis public-API benchmark shows approximately 1,681 versus 475, about 3.5×. Lie’s architectural argument is that Groq’s smaller SRAM pools make frontier-model scaling harder. Cerebras packs 44GB of SRAM per wafer, approximately 88× the 500MB per Groq 3 LPU. More local memory can reduce the number of devices needed to hold weights and the communication between them. Cerebras’s dedicated inference catalog includes Kimi, DeepSeek and GLM, alongside its wafer-scale training capability. The roadmap adds another reason for conviction: CS-4 demonstrated over 4,400 tokens/sec per user on GPT-OSS-120B. CS-5 targets up to 10,000 on medium-sized models and 5,000 on frontier models in 2027. Faster generation can support more reasoning and agent iterations within the same time budget. The market has been brutal across semiconductors: 22 of the 30 PHLX Semiconductor Index stocks sit more than 20% below their record highs. Another leg lower is possible. Timing the bottom is anyone’s guess. I view this as an opportunity to own a future powerhouse at an attractive price. The IP, demonstrated growth and established relationships give me conviction that this business can become massive. I’m more bullish than ever, volatility included. Ultrafast inference is the future.

Currently on a group call with the Macquarie analyst that just initiated coverage on $CXMT I’ll post my notes in a little bit. Will say this… if you’re selling the big 3 memory companies because the financial media says $CXMT is a threat… then you’re just as dumb as the journalists spreading this FUD. $CXMT is not a threat to Micron, Samsung or SK Hynix and I’ll share why in my notes.

UBS sees 2027 as the inflection year for $CBRS with revenue growing ~4x to $3.3B before reaching ~$23B by 2030 with 46% operating margins. The interesting part is Cerebras’s on chip memory advantage becomes way more valuable as AI agents make dozens of model calls per task with CS-5 targeting 10K tokens per second per user in 2027. And with ~$25B in contracted backlog already equivalent to nearly half of projected revenue from 2027 through 2030, shares now trade below 20x 2028 earnings.