TrendForce update on Q4 DRAM sales. Watch just the huge growth YoY on revenue from the Big3.
Samsung & SK are just raking it in w/ 43% & 25% QoQ growth respectively.
Others is CXMT + other Chinese players is up 44.5% QoQ to 3.78B.
As such, CXMT is already bigger than SMIC in revenue & growing fast. We will see its quarterly earning this year as it IPOs. It is now likely close to 7% mkt share, but capacity is much higher.
As its yield goes up & move into more higher priced chips, its share likely goes up even more.
Kris Powers, CFA retweeted
Kalsi has more alpha on rates than fed funds futures
👀
This is a big deal
The Federal Reserve just put out an incredible paper about Kalshi's data.
"Our results suggest that Kalshi markets provide a high-frequency, continuously updated, distributionally rich benchmark that is valuable to both researchers and policymakers."
federalreserve.gov/econres/f…
Kris Powers, CFA retweeted
Peter Steinberger is joining OpenAI to drive the next generation of personal agents. He is a genius with a lot of amazing ideas about the future of very smart agents interacting with each other to do very useful things for people. We expect this will quickly become core to our product offerings.
OpenClaw will live in a foundation as an open source project that OpenAI will continue to support. The future is going to be extremely multi-agent and it's important to us to support open source as part of that.
Kris Powers, CFA retweeted
Anthropic raised its 2026 revenue forecast by ~20%.
It now sees revenue up to $18B in 2025 and $55B in 2026, and lifted its optimistic 2025–2028 path by ~40%.
On that track it reaches ~$148B in 2029, while projecting $100B+ in model training costs by 2029 and a later timeline to go cash-flow positive.
Source: The Information
Taiwan electronics assembly giant Pegatron forecast triple-digit growth for its AI server business this year, led by shipments of Nvidia GB200 and GB300-based servers as well as some ASIC-based servers, media report, citing Pegatron executives at their year-end banquet. 1/2 $NVDA $TSM #Pegatron ec.ltn.com.tw/article/breaki…
1
1
206
Kris Powers, CFA retweeted
Taiwan electronics assembly giant Pegatron forecast triple-digit growth for its AI server business this year, led by shipments of Nvidia GB200 and GB300-based servers as well as some ASIC-based servers, media report, citing Pegatron executives at their year-end banquet. 1/2 $NVDA $TSM #Pegatron ec.ltn.com.tw/article/breaki…
Kris Powers, CFA retweeted
US Inflation has further dropped to 1.20% today.
Main drivers of this latest in a series of disinflationary coolings were:
- Food - mostly Eggs
- Household durables - particularly housekeeping supplies
- Alcohol & tobacco - mostly alcoholic beverages
Our number is derived by aggregating millions of real-time price data points every day to calculate a year-over-year CPI % rate. It is comparable but not identical to the survey-based official headline inflation released monthly by the BLS, which was 2.7% for December.
Kris Powers, CFA retweeted
Global AI Tracker Update
Generative AI Website Worldwide Traffic Share - Key Takeaways:
→ ChatGPT's decline stops, with share up 0.1%.
→ Gemini continues growing its share.
→ Grok surpasses DeepSeek for the first time.
🗓️ 12 Months Ago:
ChatGPT: 86.6%
Gemini: 5.3%
Perplexity: 1.9%
DeepSeek: 1.6%
Claude: 1.5%
Copilot: 1.3%
🗓️ 6 Months Ago:
ChatGPT: 78.3%
Gemini: 8.7%
DeepSeek: 4.5%
Grok: 2.6%
Perplexity: 1.6%
Claude: 1.5%
Copilot: 1.2%
🗓️ 3 Months Ago:
ChatGPT: 72.9%
Gemini: 13.3%
DeepSeek: 4.1%
Perplexity: 2.4%
Grok: 2.4%
Claude: 2.0%
Copilot: 1.2%
🗓️ 1 Month Ago:
ChatGPT: 66.8%
Gemini: 19.5%
DeepSeek: 3.8%
Grok: 3.0%
Perplexity: 2.1%
Claude: 2.0%
Copilot: 1.2%
🗓️ Today (January 16):
ChatGPT: 64.6%
Gemini: 22.0%
Grok: 3.5%
DeepSeek: 3.3%
Claude: 2.1%
Perplexity: 1.9%
Copilot: 1.1%
Kris Powers, CFA retweeted
Leading Gen AI tools' QoQ change in website visits - 2025.
Tool-by-tool breakdown in the thread below.
>>
Cloudflare CEO Matthew Prince says Googlebot sees 3.2x more of the web than OpenAI, and 4.8x more than Microsoft.
And he worries this advantage will allow Google to run away with the AI race, with no one else being able to catch them.
"For every one page that OpenAI sees, Google is seeing 3.2."
"What I worry about is, because Google has this unique access to the web that nobody else has, the game might just go to them. Because at the end of the day, whoever has the most data wins in the era of AI."
Kris Powers, CFA retweeted
a lot more hard work still to do of course, but making relentless progress...
First Global AI Tracker of 2026
Gen AI Website Worldwide Traffic Share, Key Takeaways:
→ Gemini surpassed the 20% share benchmark.
→ Grok surpasses 3% and is approaching DeepSeek.
→ ChatGPT drops below the 65% mark.
🗓️ 12 Months Ago:
ChatGPT: 86.7%
Gemini: 5.7%
Perplexity: 1.9%
Claude: 1.5%
Copilot: 1.5%
🗓️ 6 Months Ago:
ChatGPT: 78.6%
Gemini: 8.6%
DeepSeek: 4.8%
Grok: 2.1%
Perplexity: 1.6%
Claude: 1.5%
Copilot: 1.1%
🗓️ 3 Months Ago:
ChatGPT: 74.1%
Gemini: 12.9%
DeepSeek: 3.7%
Perplexity: 2.4%
Grok: 2.0%
Claude: 2.0%
Copilot: 1.2%
🗓️ 1 Month Ago:
ChatGPT: 68.0%
Gemini: 18.2%
DeepSeek: 3.9%
Grok: 2.9%
Perplexity: 2.1%
Claude: 2.0%
Copilot: 1.2%
🗓️ Today (January 2):
ChatGPT: 64.5%
Gemini: 21.5%
DeepSeek: 3.7%
Grok: 3.4%
Perplexity: 2.0%
Claude: 2.0%
Copilot: 1.1%
>>
Kris Powers, CFA retweeted
I vibecoded an open-source poker river solver over the holiday break. The code is 100% written by Codex, and I also made a version with Claude Code to compare.
Overall these tools allowed me to iterate much faster in a domain I know well. But I also felt I couldn't fully trust them. They'd make mistakes and encounter bugs, but rather than acknowledging it they'd often think it wasn't a big deal or, on occasion, just straight up try to gaslight me into thinking nothing is wrong.
In one memorable debugging session with Claude Code I asked it, as a sanity check, what the expected value would be of an "always fold" strategy when the player has $100 in the pot. It told me that according to its algorithm, the EV was -$93. When I pointed out how strange that was, hoping it would realize on its own that there's a bug, it reassured me that $93 was close to $100 so it was probably fine. (Once I prompted it to specifically consider blockers as a potential issue, it acknowledged that the algorithm indeed wasn't accounting for them properly.) Codex was not much better on this, and ran into its own set of (interestingly) distinct bugs and algorithmic mistakes that I had to carefully work through. Fortunately, I was able to work through these because I'm an expert on poker solvers, but I don't think there are many other people that could have succeeded at making this solver by using AI coding tools.
The most frustrating experience was making a GUI. After a dozen back-and-forths, neither Codex nor Claude Code were able to make the frontend I requested, though Claude Code's was at least prettier. I'm inexperienced at frontend, so perhaps what I was asking for simply wasn't possible, but if that was the case then I wish they would have *told* me it was difficult or impossible instead of repeatedly making broken implementations or things I didn't request. It highlighted to me how there's still a big difference between working with a human teammate and working with an AI.
After the initial implementations were complete and debugged, I asked Codex and Claude Code to create optimized C++ versions. On this, Codex did surprisingly well. Its C++ version was 6x faster than Claude Code's (even after multiple iterations of prompting for further optimizations). Codex's optimizations still weren't as good as what I could make, but then again I spent 6 years of PhD making poker bots. Overall, I thought Codex did an impressive job on this.
My final request was asking the AIs if they could come up with novel algorithms that could solve NLTH rivers even faster. Neither succeeded at this, which was not surprising. LLMs are getting better quickly, but developing novel algorithms for this sort of thing is a months-long research project for a human expert. LLMs aren't at that level yet.
Very interesting.
DRAM makers are now requiring customers to sign LTAs that do not allow cancellations or returns.
[Commentary on DRAMeXchange December Server DRAM Fixed Prices & Industry Trends]
- December Pricing & Additional Orders: Negotiations for December Server DRAM fixed prices are proceeding at elevated levels, with additional orders for December delivery becoming a key topic. Major customers include US CSPs, Chinese CSPs, and OEMs. As supplier inventories are depleted, additional orders can only be fulfilled by rescheduling delivery timelines with existing customers. Consequently, actual transaction prices in December surged, rising 88–93% compared to the 3Q25 average fixed price. Furthermore, the 4Q25 fixed price is expanding to a level 53–58% higher QoQ.
- 1Q26 Outlook & Supplier Stance: From December, the negotiation focus is gradually shifting to 1Q26 contracts. Given the current tight supply-demand conditions, suppliers are expected to propose initial quotes reflecting a 60–70% increase, exceeding the levels seen in December's add-on orders. Suppliers maintain a bullish outlook for continued quarterly price hikes throughout 2026. Anticipating further increases in future quarters, they are currently reluctant to sign multi-year lock-in contracts with CSPs.
- 2026 Price Forecast: Considering the above, the growth rate for Server DRAM fixed prices in 2026 is projected as follows: 1Q (+60–65%) / 2Q (+10–15%) / 3Q (+3–8%) / 4Q (+0–5%). This represents an upward revision from last month's estimates, and further upward adjustments cannot be ruled out.
- CSP Financials & Demand: CSPs assess that cash flows generated from existing businesses can sufficiently cover CapEx for large-scale infrastructure build-outs, with a focus on realizing monetization from inference AI. Therefore, they are not significantly pushing back against DRAM price hikes, accepting them to secure stable volumes. Analysis suggests that considering their margin structures, customers still have room to absorb these price increases.
- Market Monitoring: Market participants are closely watching these purchasing strategies to determine when the peak of the price (P) cycle will occur. OEMs are expected to maintain margin headroom until at least 1H26, while CSPs are projected to be capable of absorbing potential supply gluts in the market.
- HBM3E Impact & Supply Allocation: Entering December, GPU/ASIC manufacturers retracted previous expectations for a decline in HBM3E purchase prices in 2026, as suppliers successfully induced price hikes against the backdrop of shortages. Consequently, suppliers are reducing legacy commodity DRAM supply to focus on HBM3E. Amid this trend, adjustments to supply and procurement volumes for CSPs and OEMs continue within 2026 LTAs (Long-Term Agreements). Some OEMs and Chinese CSPs with relatively lower transaction stability have reportedly received notices of supply cuts. Overall, while Server DRAM supply volumes are being revised upward, PC DRAM remains an application area where suppliers are concentrating on production cuts.
- Customer Prioritization: Supplier allocation decisions are based on the customer's actual demand levels, past transaction history, and partnership evaluation. With US CSPs showing rising transaction prices for Server DRAM and HBM while actively securing 1H26 deliveries, the possibility cannot be excluded that suppliers may cut Server DRAM supply for lower-priority customers. This increases the likelihood of intensifying tight supply conditions for OEMs and Chinese CSPs.
- AI-Driven Demand Growth: As inference AI applications like Gemini and Copilot drive demand for commodity servers, CSPs (led by MSFT and Google) are aggressively increasing server assembly orders to ODMs. Accordingly, the 2026 server shipment growth forecast has been revised upward from 9% to 13% YoY. Additionally, the expanded adoption of multi-core CPUs is increasing procurement of high-capacity RDIMMs, pushing Server DRAM content per box up 20% YoY. Total Server DRAM bit demand growth is expected to reach 39%, significantly exceeding the previous forecast of 23%.
- Contract Terms (NCNR): Suppliers are negotiating LTA contracts that include NCNR (Non-Cancellable, Non-Returnable) clauses with major customers, aiming to secure long-term CapEx visibility through 2028 based on these agreements. However, despite these contracts, securing sufficient supply in a timely manner for 2026 is expected to remain challenging.
(Source: DRAMeXchange)
Kris Powers, CFA retweeted
H200 2026 Supply Math: N4 Wafer Starts, CoWoS-S WPM Allocation, and HBM3E Spend
Nvidia currently has 700K H200s in inventory, but China has ordered more than 2M H200s. So TSMC will have to fab 1.3M H200s in 2026. Nvidia will generate over $54 Billion in revenue (@ $27k per H200 unit)
At 55 dies per wafer (85% yield), TSMC needs to dedicate nearly 24,000 N4 wafer starts. Allocating 3,000 WPM N4 capacity over 8 months should be enough. It will generate nearly $450M-$500M in revenue for TSMC (@ $18k-$20k per wafer)
CoWoS is the real constraint
TSMC has 80k WPM of CoWoS capacity rn. 35% of this is reserved for CoWoS-S, so nearly 28k WPM.
One CoWoS wafer can yield 28 H100/H200. TSMC needs a total of 46k CoWoS wafers to package 1.3M H200s.
They need to dedicate 3.8k WPM (12 months) to 5.75k WPM (8 months) to fulfil the H200 orders. Each CoWoS wafer costs $12k.
So, they will generate 46k * $12k = $552 Million from packaging
Total TSMC Revenue= $550M + $500M = $1.05 Billion
HBM3E demand
HBM3E price is $15 per GB in 2025. A decline to $10-$12 per GB was expected in 2026, but it won't happen now due to China's demand.
So, for 1.3M GPU each equipped with 141GB of HBM3E, we get: 141*1.3M*15 = $2.75 Billion (183M GB) in HBM revenue for the memory makers
Nvidia Revenue = $54 Billion
TSMC Revenue = $1.05 Billion
Memory Revenue = $2.75 Billion
Kris Powers, CFA retweeted
Every ASIC is SRAM based.
Literally levels of SRAM from compute to memory.
It's what L1, L2, L3 etc is. It's SRAM.
Stop treating it like some sort of 'new' mystic paradigm. It's always been about how much you put close to the compute for the performance/die-area tradeoff.
Kris Powers, CFA retweeted
I was inspired by this so I wanted to see if Claude Code can get into my Lutron home automation system.
- it found my Lutron controllers on the local wifi network
- checked for open ports, connected, got some metadata and identified the devices and their firmware
- searched the internet, found the pdf for my system
- instructed me on what button to press to pair and get the certificates
- it connected to the system and found all the home devices (lights, shades, HVAC temperature control, motion sensors etc.)
- it turned on and off my kitchen lights to check that things are working (lol!)
I am now vibe coding the home automation master command center, the potential is 🔥.And I'm throwing away the crappy, janky, slow Lutron iOS app I've been using so far. Insanely fun :D :D
Kris Powers, CFA retweeted
AI on a wicked race to the bottom!
Grok 4: $0.28/M tokens
Gemini 3 Flash: $0.50/M tokens
Gemini 2.5 Flash: $0.30/M tokens
Gemini 2.5 Flash-Lite: $0.10/M tokens
GPT-4 in ’23: $37.50!
Look out below, and let’s go! ❤️
Kris Powers, CFA retweeted
Nvidia is buying Groq for two reasons imo.
1) Inference is disaggregating into prefill and decode. SRAM architectures have unique advantages in decode for workloads where performance is primarily a function of memory bandwidth. Rubin CPX, Rubin and the putative “Rubin SRAM” variant derived from Groq should give Nvidia the ability to mix and match chips to create the optimal balance of performance vs. cost for each workload. Rubin CPX is optimized for massive context windows during prefill as a result of super high memory capacity with its relatively low bandwidth GDDR DRAM. Rubin is the workhorse for training and high density, batched inference workloads with its HBM DRAM striking a balance between memory bandwidth and capacity. The Groq-derived "Rubin SRAM" is optimized for ultra-low latency agentic reasoning inference workloads as a result of SRAM’s extremely high memory bandwidth at the cost of lower memory capacity. In the latter case, either CPX or the normal Rubin will likely be used for prefill.
2) It has been clear for a long time that SRAM architectures can hit token per second metrics much higher than GPUs, TPUs or any ASIC that we have yet seen. Extremely low latency per individual user at the expense of throughput per dollar. It was less clear 18 months ago whether end users were willing to pay for this speed (SRAM more expensive per token due to much smaller batch sizes). It is now abundantly clear from Cerebras and Groq’s recent results that users are willing to pay for speed.
Increases my confidence that all ASICs except TPU, AI5 and Trainium will eventually be canceled. Good luck competing with the 3 Rubin variants and multiple associated networking chips. Although it does sound like OpenAI’s ASIC will be surprisingly good (much better than the Meta and Microsoft ASICs).
Let’s see what AMD does. Intel already moving in this direction (they have a prefill optimized SKU and purchased SambaNova, which was the weakest SRAM competitor). Kinda funny that Meta bought Rivos.
And Cerebras, where I am biased, is now in a very interesting and highly strategic position as the last (per public knowledge) independent SRAM player that was ahead of Groq on all public benchmarks. Groq’s “many chip” rack architecture, however, was much easier to integrate with Nvidia’s networking stack and perhaps even within a single rack while Cerebras’s WSE almost has to be an independent rack.








































