sigma capitalist

$MRVL held a webcast today on its acquisition of Lumentum. On slide 13 they show an NPO module with a DSP in it (pictured below) Our channel checks indicate that Marvell is working on a TX-only DSP for NPO. The transmit side is retimed and the receive side stays linear, so it is not a fully retimed DSP. That adds about 30 to 35 ns of latency, against roughly 90 ns for a fully retimed DSP. The point of NPO was to sit close enough to the ASIC to go without a DSP. The electrical path from the ASIC to the optics was meant to be short enough to skip the DSP. It looks like the electrical losses on the host side and the need for stable scale-up links are enough to warrant a retimer on the transmit side after all. It is still early, but adding a DSP would reduce NPO’s power and latency advantages over pluggables. It would also give Marvell another DSP socket inside a design that was originally intended to not work with one.

Whenever news break about a rumored M&A, and if it's indeed true, the actual formal announcement typically follows within 48 hours to lock in deal terms But there's still nothing on $MRVL - $LITE Reasons could be: > both sides have not or cannot agree on terms (possibly due to the sell-off in both stocks, and the overall tech/semi sector) > financing (mezz/bridge, if any) and deal structure issues (stock exchange ratio, cash terms, break fee, go-shop, etc) > new risk factors that could have arisen - antitrust, regulatory, etc I also think the stock premium that Marvell could offer for Lumentum, if they proceed, may not be as high as what many people think. Definitely not 30s, maybe 20s, or could be as low as mid-teens The agreed premium will be an indicator of how Lumentum board/senior execs think about the future prospects of their business, maybe even the overall optics/photonics sector

$CBRS Cerebras - bear case thesis Today (June 25) is the day when 30.2M shares held by existing directors, officers and non-employee shareholders (i.e. pre-IPO investors) are unlocked. The unlock starts at 6:00AM ET During the IPO, Cerebras issued 34.5M shares. The first release of shares shall result in an increase of ~87.5% in free float, from 34.5M shares to 64.7M shares, assuming all the unlocked shares are converted to Class A shares for trading In dollar terms, at last close ($182.41), this represents ~$5.5B Bear case thesis: 1 – the announcement by OpenAI to develop a custom silicon inference accelerator with Broadcom (codenamed Jalapeño) is a potential negative for Cerebras Reason: under the terms of the OpenAI Master Relationship Agreement (MRA) with Cerebras, OpenAI has two deployment options – i.e. to receive Committed Compute via Cerebras Cloud, or via the purchase and deployment of Cerebras systems in their own data centers (Note: Committed Capacity refers to the 750MW of AI Inference compute capacity that will be deployed in tranches from this year to 2028, with each tranche having a term of 3 or 4 years, extendable to 5 years. OpenAI also has the option to purchase Additional Capacity of 1.25GW for deployment in tranches by end of 2030) OpenAI’s choice of deployment will impact Cerebras’ revenue mix (hardware versus cloud and other services) as well as segment/overall gross margins If OpenAI elects to receive compute via Cerebras Cloud, impact on gross margin is overall negative because such revenue attributable to OpenAI is recognized as pass-through revenue (booked under cloud and other services) and reported on a gross basis with a 3% mark-up – i.e., pass-through revenue = pass-through costs + 3% margin Pass-through costs comprise of: fixed monthly data center rental costs, leasehold improvements, security and other variable monthly costs (e.g. power / utility) Revenue recognized from OpenAI is also subject to GAAP non-cash amortization due to the warrants issued to OpenAI Cerebras currently assumes the initial 250MW of committed capacity to be delivered to OpenAI through Cerebras Cloud Assuming if OpenAI successfully develops and deploys its own custom inference silicon, it is highly likely that they may choose to receive compute through Cerebras Cloud for the fulfilment of the Remaining Performance Obligations (RPO). In this scenario, Cerebras’ overall margin will deteriorate further 2 – Terms of their agreement with AWS are quite scant, but from their recent 10-Q, they include an “initial multi-year lease of the Cerebras systems” and “options for future procurement of products and services”, with “minimum manufacturing capacity guarantees” This means that instead of purchasing Cerebras systems outright to be deployed together with Trainium for disaggregated AI inference workloads (Trainium for pre-fill, CS-3 for decode), AWS will initially lease these systems and deploy them in their own data centers This arrangement compounds the decline in hardware sales over the next few quarters, as guided by Cerebras’ CFO, into 2027 when revenue contribution from AWS starts materializing (Note: during the 1Q earnings call, Cerebras' management declined to answer an analyst's question on the scope of the AWS agreement across the compute spectrum and scale) Apart from finding data center space/capacity to fulfil its obligations to OpenAI, Cerebras has another problem on hand – which is to find new customers to purchase its hardware/systems 3 – Customer concentration / financial accounting G42 and MBZUAI, both related entities, contributed about 74% of Cerebras’ 1Q revenue, and 54% of accounts receivable While its partnership with OpenAI is supposed to mitigate customer concentration risk, Cerebras currently presents core gross profit as GAAP gross profit adjusted to exclude: i) Pass-through revenue and related data center costs ii) Non-cash amortization of customer warrant assets iii) Stock-based compensation expense Given that items i) and ii) are largely attributable to OpenAI (and subsequently AWS for the warrants), it’s unclear how they plan to present their financials if majority of OpenAI’s remaining deployment is going to be through Cerebras Cloud and with such revenue to be booked as pass-through revenue 4 – Stock overhang Shares held by pre-IPO investors and employees/directors/officers are scheduled to be unlocked starting today, with a final release on the earlier of the 2nd trading day after Q3 26 or 180 days after IPO pricing (around November 9, 2026) A total of >210M shares will be unlocked from now until early November = $38B+ (on last closing price) This has not yet factored in the shares underlying the warrants issued to OpenAI and AWS totalling ~36.1M Class N shares, freely convertible to Class A common stock once vested and exercised

ANTHROPIC Q3 2026 FINANCIAL HIGHLIGHTS: - Revenue: $17.0B, +1,127% YoY & 47% QoQ - Oper Loss: $567M (vs. $974M profit in Q2) - Net Loss: $23.3B, vs. $12.1B YoY - Adj. Oper Income: $2.5B - Adj. Oper Margin: 15%, down from 19% in Q2 - 9M Revenue: $33.2B Anthropic attributed its return to an operating loss partly to $5B in higher-cost short-term compute and a $2B noncash donation-matching charge. Its nine-month net loss reached $104B, largely reflecting $94.3B in noncash charges from the rising value of its convertible notes. Anthropic also disclosed approximately $895.4B in aggregate purchase and lease commitments as of and after Sept. 30, largely tied to AI compute infrastructure.

A network of brokers engineered trades to mint billions for Indonesia’s tycoons – until MSCI triggered a crackdown bloomberg.com/news/features/…

last month i was begging gpu providers for b300s at $7.9/hr now they're begging me to take them at $4.6/hr, no commitment something big is happening soon. ai gave us all its dessert and now it's time for the veggies

The cost of serving Grok Bot was unsustainably high relative to revenue

Quick thread on AI and productivity: Lots of people say this chart shows AI is now showing up in the productivity data. But I made this chart to argue the opposite: a trend that pre-dates AI and only shows up in the US doesn’t look like AI. Here’s how I’m thinking about the data and underlying dynamics:

Did a quick sweep of my X wall to see what's trending and saw a few accounts pumping $RDDT Reddit I don't know if people realize that Reddit *was* great as a data scraping ground for AI labs (initially, then with licensing agreements) due to its pool of human-generated content which these labs require to train their models/NLP capabilities However, the situation has since evolved. The advanced/frontier AI labs are now more focused on highly curated content - and are willing to pay for specialized domain knowledge/data work. Platforms like Mercor are thriving because of this For AI startups, they can now bootstrap training/fine-tuning using synthetic data (like TML's Inkling using Chinese open models) Those who say, well, there're data integrity issues with AI training on AI But the counterargument is, the internet is increasingly filled with AI-generated content and what's more, we now have agent swarms generating/iterating data loops Not only is signal-to-noise ratio shifting rapidly, we'll soon be unable to discern between authentic human-generated content vs AI ones My view is public forums like Reddit are soon going to be what they're initially set out to be - mostly for fun and entertainment purposes. Somewhere you doom-scroll mindlessly after a long work day lol Similarly for Discord. Arguably, they are better positioned given their community vibes with dedicated spaces/whatnot but from what I understand, they are facing real monetization difficulties. I believe they have been/are trying to IPO but I think it's gonna be a tough sell

That is appalling. That means the young couple who died won’t even get any compensation. straitstimes.com/singapore/c…