With Anthropic’s Mythos model and OpenAI’s GPT 5.5, we could be looking at the new normal for LLM releases by capacity-constrained LLM builders.

In the case of Mythos, it was released weeks ago to a preferred group of customers.

In the case of GPT 5.5 it is being released in stages through the summer, and its cyber “module” (which carries the 5.4 label), is initially limited in scope via its TAC circle of trust program, which reads like Mythos’ marketing, which is to say it reads like nonsense.

My sense is that given the capacity constraints of Anthropic in particular and OpenAI, these new GPU-hungry models are too expensive for the masses and therefore are being rolled out in a phased approach so that model costs may come down as GPU capacity is on-boarded.

To that end, if it is true that xAI has only utilized 11% of its NVIDIA chip cluster - granted it loaned some to Cursor and is in-process of acquiring Cursor - why not rent GPU capacity to Anthropic? Perhaps xAI feels that Anthropic is too much of a direct LLM competitor and would rather see Anthropic starve for capital and GPUs.

It is not certain that xAI will ever develop the best coding model, therefore I believe that xAI should jump into the chip rental market and compete with the likes of CoreWeave (CRWV) and others.

I am looking forward to trying Anthropic’s Mythos for coding. Opus 4.7 was definitely a step back from Opus 4.6 on the coding front. I’m still using Opus 4.6 and OpenAI’s Codex 5.5.

I want to use Mythos before I call bogus to Boris Cherny’s (creator of Claude Code), claim that coding is solved. It isn’t.

Meanwhile, Google is the one LLM company that isn’t starved for capital nor chips. It can slowly chip away at improving Gemini’s coding and agentic capabilities as its purpose-built TPU chips continue to improve.

Several housekeeping items before I get to AI pricing:

1.) Our TEK2day article archives are now free to access.

2.) The Kilby earnings transcript “debrief” product that I mentioned yesterday is now publicly available HERE as we completed testing.

3.) The next in-depth article that I publish will likely cover the state of AI, specifically the “harness” and “agentic” layers. The AI industry has invested heavily in infrastructure from data centers (AMZN, Anthropic, GOOG, META, MSFT, OpenAI, ORCL, etc.) to DC connectivity suppliers (CSCO, CIEN, ADTN, Huawei, etc.) to silicon producers (NVDA, AMD, AVGO, INTC, etc.) to AI governance and analytics (AMZN, GOOGL, MSFT, ORCL), to LLMs (AMZN/Kiro, Anthropic, Google/Gemini, IBM, META, Mistral, MSFT, OpenAI, etc.), but the real innovation is happening at the harness and agentic layers where companies are wrapping intelligence around the LLMs - and running fast to do so as LLMs rapidly enhance their capabilities.

AI has changed the way software companies price their various offerings including applications and AI agents.

I touched on this subject a couple of weeks ago.

The reason for this pricing evolution is because software companies that build on top of Anthropic (as we have done at Kilby), OpenAI, and Gemini are charged consumption fees by the model companies for API usage. Software companies in turn pass the API fee cost on to users either via a straight consumption-based model or hybrid model (hybrid = baseline fee plus consumption fee above a certain threshold), whereas fewer companies are charging straight subscriptions.

This trend will continue as more Software companies bake AI into their B2B offerings. I do not believe these pricing changes will bleed into Consumer offerings other than for consumers who are heavy users of AI-imbued products.

The below graphic comes from a talk that the payments company Stripe published two days ago. I have included that talk as well.

We will push this new product out to Kilby users in a few days.

META is up 30% over the past month which doesn’t make sense to me.

META has failed at crypto, has failed at the metaverse, and will fail at building a frontier LLM. META does not have an AI-first culture. That’s the main reason why META will fail at building a leading frontier LLM. Further, China just blocked META’s acquisition of Manus.

I wish Zuckerberg stuck to the original plan of focusing on opensource LLMs, that plan had merit.

If META hadn’t cut heads last week, I would say the stock would be off 30% after Wednesday’s EPS call, but the RIF may buy META some air cover. I can’t imagine that the Ad outlook will be strong.

Treasury Secretary Bessent said that UAE and others in the region approached the U.S. about executing currency swaps, meaning the region has an appetite for Dollars, most likely needed to pay down Dollar-denominated debt. So if UAE and Saudi own approximately $280 billion in Treasuries combined and each needs Dollars to pay down debt, that suggests there is pressure to sell at least some of the $280 billion Treasury position.

How many other countries are sucking wind and considering selling Treasuries? Yields have to go higher. I’m sorry, whomever is in love with the 10YR at a 4-5% yield is a fool.

Heck, at this point I am hoping China will exert heavy pressure on Iran to open the Strait of Hormuz soon. If the U.S. starts dropping bombs on Iran, and if Iran in turn drops bombs on the UAE and Israel, oil could shoot from $100 to $200 in a week’s time or less. If the U.S. hammers Iranian oil production, it’s not as though that supply could be built up in a month. $10 gas at the pump does not sound like fun.

The U.S. is playing with fire, I’m not sure why. All we’ve done is drive the price of oil 50%+ higher while growing the fiscal deficit ($1.2 trillion YTD), which has only accelerated the Dollar’s decline. At this rate the USD won’t be the global reserve currency for much longer.

Pay as you go pricing models are displacing subscription models because the model providers (OpenAI, Anthropic) charge corporate customers for usage via their API.

As a result, software companies that build software applications on top of the model providers are doing one of two things - moving toward pay-as-you-go pricing models - which keeps revenues aligned with costs, or they are inflating their subscription fees to protect against customer usage spikes.

To the latter point, let’s say for example that for 8 months of the year a given customer consumes the equivalent of $100 worth of usage for an app that sits on top of Anthropic, but in the other 4 months the customer consumes $300, $500, $250 and $700 worth of usage. You can see how a subscription model would break under that scenario. The app developer’s income statement would take a meaningful hit unless the developer over-priced its app at let’s say $1,000/month to protect against downside risk.

Alternatively, usage-based models align customer usage with vendor pricing. Under a usage based model, the app developer can price its applications at a spread over the usage fees it is charged by the model provider (developers buy credits from LLM providers for API usage and in turn sell usage credits at a markup over cost).

Here is Anthropic’s pricing sheet: https://platform.claude.com/docs/en/about-claude/pricing

A developer would charge a spread over these usage costs. This is fairly common today and will become the dominant Enterprise software pricing model for AI-infused applications.

As I mentioned earlier today I am really enjoying coding with Codex 5.5 in terminal. I always use AI coding tools in terminal as operating locally is much faster than operating in the cloud, which is why I built a terminal option for Kilby.

OpenAI has stated that there are more than 3 million Codex users out of more than 900 million total GPT users. I am not sure how many Codex users operate in terminal.

OpenAI has always been best-in-class in terms of the model’s personality. From a user experience standpoint, OpenAI is very generous with usage, especially in comparison to Anthropic’s throttling. Now for the first time in a long time OpenAI’s coding capability are also best-in-class, at least for now before Anthropic’s Mythos is released.

One of the Mythos rumors I heard a few weeks ago was that the model was too expensive to release, which is likely the reason Anthropic acquired more GPUs several weeks ago in an effort to increase capacity and bring down pricing. I have not spoken with anyone who has used Mythos to build anything.

It is also interesting that OpenAI reduced the cost of its top pricing tier from $200/month to $100/month. OpenAI has more GPUs than Anthropic. I believe that OpenAI has a larger balance sheet than Anthropic and it feels like OpenAI is trying to put the screws to Anthropic in advance of a Q3 or Q4 IPO.

I would try to punch out the IPO now if I were advising either company. I don’t know that either CEO or Board appreciates how fast things could go sideways if the Iran War persists or escalates. I fully expect it to persist because the DoD in service to its corporate masters (Lockheed, Boeing, Northrop, RTX), needed something to replace the revenue stream generated by Afghanistan. Iran is that “something”.

I am using OpenAI’s Codex 5.5 model to debug code now. I can say I like it better than both Anthropic’s Opus 4.6 model and certainly better than Opus 4.7, which feels like a downtick as it relates to coding. Codex 5.5 is OpenAI’s vaunted “spud” model. We shall see how Anthropic’s Mythos model stacks up when it is released.

2026 should be interesting insofar as the Anthropic vs. OpenAI battle is concerned for the reason we wrote about a couple of months ago in that OpenAI acquired GPUs more aggressively than Anthropic in 2025. This is part of the reason why Anthropic throttles usage, the company is compute capacity constrained.

Meanwhile, Sergey Brin has been back at Google DeepMind for I’m not sure how long trying to improve its models. The issue there is that DeepMind does great research, but the lab struggles to commercialize its research. I still believe Sundar Pichai is not the right person to lead the company. Whomever leads DeepMind when the dust settles ought to be GOOG CEO.

So, for the moment my preferred coding model is OpenAI Codex 5.5.

The phrase that comes to mind when I think of Tim Cook is “opportunity cost”.

What Apple (AAPL) could have been had it focused on technology beyond a thinner battery over the past decade-plus. The Technology industry is moving too fast to not have a Product/Engineering person at the helm. Tim Cook will step down as CEO on Sept. 1st. CEO-in-waiting John Ternus is Apple VP of Hardware Engineering.

There was a time when Tim Cook - Apple’s former COO - considered his successor - former Apple COO Jeff Williams for the CEO role. What a mistake that would have been to have two supply chain guys as consecutive AAPL CEOs. I joked the other day that Tim Cook may have been the only mistake of consequence that Steve Jobs ever made as Apple CEO.

Here’s hoping that John Ternus can help right Apple’s ship on the AI side. Apple ought to have its own fleet of top tier models optimized for the iPhone, but that is easier said than done and would require a culture change. No need to wait until September 1st for the formal transition date. Ternus ought to lead the AI charge now.

Kilby customer note: Claude Opus 4.7 is now a model option.

It’s clear to me that Google / DeepMind do not care about building the best-in-class LLM. Rather, Google is building a language model that can plug into Google’s various products including Search, Workspace, Android, Cloud, YouTube, Pixel, advertising and so on. Gemini very much feels like a feature, not a product.

I get the sense that Demis Hassabis - CEO of Google DeepMind - is 100% focused on AGI research, but not on developing the best commercial LLM. Anthropic in my view has the most focused mission - to build the best model to service Enterprise API users.

I know Demis sent the below chart around, but heavy users of Claude aren’t hanging around the Website, we’re in terminal.

If Google is simply trying to enhance its product family via Gemini such that it may justify price increases, it should consider making Gemini free, or make the $20 Gemini plan the usage equivalent of Anthropic’s $100 monthly plan. If you’re not trying to build the best model, don’t charge as if you were. Opensourcing Gemini could drive the adoption Google hopes to achieve.

Separately, some people have asked me what is Kilby? I think of Kilby as an intelligent AI assistant that has agentic capability. Kilby is both a general purpose AI assistant and an Investment Management-focused AI assistant given that Kilby has built some 40 tools that are Investment Management-specific, not to mention various skills and enhanced memory techniques that enable Kilby to remember user-specific context such as watchlists, holdings and more as it relates to Investment Professionals.

AI industry professionals consider Kilby a “harness”. Kilby does not directly compete with LLMs. Kilby’s intellectual property wraps around the LLM (Kilby is built around Anthropic’s Claude family of models and charges a small premium to Anthropic’s usage fees). Kilby is therefore additive to the LLM, making it more useful for specific use cases and industry verticals. Anthropic is a Kilby partner, not a competitor.

My AI assistant knows my portfolio holdings, but hasn’t yet been trained on the weightings. It will become even smarter when we soon make it automatically aware of portfolio holdings weightings.

My Prompt:

What does the re-closure of the Strait of Hormuz combined with NFLX's guidance miss mean for my portfolio?

Note: Kilby has my mock portfolio holdings which include A, AAPL, ABNB, ACN, ADBE, ADSK, AMAT, AMD, AMZN, ANET, AVGO, CSGP, GOOGL, META, MSFT, NFLX, NVDA, ORCL

Kilby response:

Good — here’s the picture across both events and what they mean for your book.

---

The Hormuz Re-closure

Iran announced Saturday it was closing the Strait again, citing “repeated breaches of trust” as the US maintained a naval blockade on Iranian ports.

The standoff escalated as Iran reversed its reopening and fired on ships attempting to pass — the Revolutionary Guard warned that approaching the strait “will be considered as cooperation with the enemy.”

Trump said US representatives are traveling to Islamabad for a new round of negotiations with Iran. So it’s open/close/open/close — a negotiation weapon, not a permanent blockade, but still functionally closed right now.

The closure of the sea lane connecting the Persian Gulf to global energy markets has triggered the largest oil supply disruption in history.

Before the conflict, roughly 25% of the world’s seaborne oil trade and 20% of global LNG passed through the strait.

What this means for your portfolio: Your book is almost entirely tech and communication services — you have essentially zero direct energy exposure, which is the right place to be in a strait closure. The transmission mechanisms that matter for your holdings are:

• Energy-driven inflation → rates staying higher → multiple compression. This is the real risk channel. If oil spikes keep CPI elevated, the Fed stays on hold or tightens, and growth-multiple tech names (NVDA, ADBE, MSFT, AMZN, GOOGL) are the most sensitive. Your semiconductor names in particular carry high multiples that don’t love a prolonged high-rate environment.

• Consumer discretionary pressure → NFLX, ABNB. Higher energy costs compress household budgets. The marginal streaming subscriber and discretionary travel consumer feel it first.

• Supply chain and capex cost inflation → AMAT, AVGO, AMD. Semiconductor capital equipment and chip manufacturing have meaningful energy and chemical input costs. Modest effect, but real.

• AMZN logistics. Fuel costs are a real P&L line for Amazon’s fulfillment network. Not existential, but margin-dilutive at the margin.

The bull case here is that Iran is using the closure as a negotiating chip, and for Iran it is perhaps its most powerful weapon, threatening the world economy and inflicting political pain on Trump. That creates strong incentive for a deal — this is probably resolved before it becomes a structural oil shock. Treat it as elevated tail risk, not base case.

---

Netflix Q1 2026 Earnings

Netflix beat Wall Street expectations for Q1 revenue, but the company announced Reed Hastings — co-founder and current chairman — would exit the board in June when his term expires. Shares fell 9% in extended trading.

The market’s reaction wasn’t really about the revenue beat.
Netflix said it expects Q2 revenue to increase 13% and warned that content spending would be front-loaded — it expects Q2 to have the highest year-over-year content amortization growth rate in 2026, before coming down in the second half. That’s a near-term margin headwind the street didn’t love.

Netflix maintained its previous full-year guidance of revenue between $50.7 billion and $51.7 billion. So the full-year isn’t broken — this is a Q2 cost-timing story plus a governance headline (Hastings departure) creating a sentiment reset.

Portfolio read: NFLX is a single-name hit in your book, not a sector signal. The underlying business — ad sales expanding rapidly and ad revenue projected to double to $3 billion in 2026
— is intact. The -9% move is likely a buying opportunity if you’re long-term, but the near-term setup is sloppy: heavy Q2 content spend, a governance overhang from Hastings’ exit, and now a macro risk-off backdrop from Hormuz. That’s three headwinds converging on one name at once.

---

Combined Portfolio Read

The two events aren’t really connected operationally, but they converge on the same macro signal: risk-off, rate-sensitive, multiple-compression pressure on growth tech. Your book is concentrated in exactly those names.

Three things to think about:
1. Do you want to trim anything near multi-year highs? NVDA, META, and GOOGL have run hard. A sustained energy shock that keeps the Fed on hold is the scenario where you most want dry powder.
2. NFLX at -9% is probably an overreaction on fundamentals — but wait for the dust to settle. Q2 guidance is the real read; H2 is the story. Don’t chase the dip today with Hormuz unresolved.
3. Watch the 10-year. If DGS10 breaks above 4.7-4.8% on energy-driven inflation fears, that’s your sell signal across the board. That’s the macro trip wire for the whole book.

The fact that AI benchmarks are gamed (obvious if you follow the model releases), is secondary to how easy it is to game the benchmarks.

Great detail in their article: HERE.

I have yet to use Claude Opus 4.7 (The model that Anthropic released yesterday), as I like to operate in terminal and I don’t see Opus 4.7 as a model option in Claude Code. Opus 4.6 is back to normal now that 4.7 is released, meaning less throttling. I used Opus 4.6 with the 1M context window all day yesterday. Models don’t quite yet reason like humans, but are getting better. By the way, 4.7 is not Mythos.

Ad: Turn PDF tables into Excel tables with Kilby:

I do expect Anthropic Mythos and GPT 5.4 Cyber to be an improvement over current models. Yet, there is a healthy dose of marketing hype in all of the recent press around each, especially Mythos. Call me skeptical given that IPOs loom for each company.

I mean, Treasury secretary Scott Bessent warning us about Mythos? Please. Ye know not what ye speak. Is the Federal Government getting warrants on the IPO? Is the hype part of Anthropic’s Govt. contract? Are the early “Project Glasswing” customers getting a discount in exchange for saying nice things?

I just don’t believe that chip architecture has sufficiently evolved to the point where LLMs can reason at the level I would like to see, which would be to exercise common sense around problem solving when writing code. The models still have blind spots to where humans must be involved in the software coding process - especially as it relates to reviewing code.

The good news is once I have Mythos for a week or two, I’ll know what’s real and what’s hype.

My strong sense is that Enterprise AI investment will continue even if Consumer AI spend slows.

The foregone cost savings are too great for companies to not deploy AI - at a minimum across the Software Engineering (SWE) function. Token spend increases will be more than offset by human capital spend decreases. Therefore, GPU production will continue.

My sense is also that CEOs are not close enough to LLMs to fully grasp what the ROI potential is across their companies. I am also not sure how many CTOs fully grasp LLMs’ potential to drive automation. You have to get in the weeds with LLMs to understand what each model is capable of, with which harnesses and tool sets etc.

My view is that there are huge cost saves to be had across Corporate Finance, Marketing, Customer service.. you name it.

As it relates to GPUs - chip companies can’t produce enough. I wonder if Tesla will acquire a piece of Intel? Tesla’s Terafab will take some time to pump out silicon at scale, INTC could help.

Last night we rolled out a new tool which turns Word docs and PDFs into markdown files, which are more quickly ingested and processed by LLMs. It’s really fast for example when you deploy a local instance of Kilby on a series of EPS call transcripts in markdown format.

You will see more services like Obsidian pop up that store your files locally in markdown format so that tools like Kilby can zip through them.

Perhaps I should know better than to complain...

• Did anyone notice last month (March 19th report date), that Accenture’s (ACN) Dollar bookings were only up 1% Y-o-Y despite a weaker Dollar? Accenture operates at the tip of the spear, Technology investors should be concerned.

• Much of ACN’s bookings weakness was a function of a soft economy. However, some of it is AI-related - LLMs make it much easier to do the software translation work that occurs when a customer switches software vendors for a particular product. Customers often hire consultants to do this translation work. LLMs can do much of that work now, in fact they can do all of it with some supervision.

• AI. Speaking of AI - software coding is the easiest use case to automate because it is 100% verifiable. Not so with most other white collar tasks, although the Corporate Finance function comes to mind. The combination of LLMs and deterministic software automations can eliminate much of the manual effort around book keeping. That said, the uptake on other white collar tasks won’t be as rapid as software development. Therefore revenue growth has to slow for Anthropic, OpenAI and Gemini beginning now.

• Oil. Physical oil hit $150 over the weekend. These elevated prices will increase the cost of everything from fertilizer to food in the coming months regardless of the futures price of oil. Real world prices will move higher. The 10YR should be much higher than 4.3% because prices will move higher during 2026.

• GDP. M1 grew 4.8% in February and is likely above 5.0% now. Nominal GDP was 5.4% in Q4 2025 and has to be lower now, maybe 1-2% given that prices continue to move higher as general economic activity/unit sales slow. Assume 2% Nominal GDP growth less 5% M1 growth = -3%, or recession by my math. That’s the more accurate way to calculate real GDP, not using the suppressed CPI measure.

  ---

I had mentioned Vercel back in early February as likely to go public in 2026. It appears the company is prepping for an IPO.

Stay on top of Technology News at T2D Pulse (free publication).

Anthropic miscalculated in 2025 as it did not buy enough compute. The CoreWeave (CRWV) deal is too little too late. The throttling is extreme at the moment as Anthropic is clearly prepping the release of Mythos despite what it says publicly.

It will be interesting to see if OpenAI can use its compute advantage to close the coding gap (it’s not a severe gap) between Claude and Codex. My guess is that both Claude Mythos and OpenAI’s “superapp” will arrive in June.

I would love to see Google push Gemini harder as Google has the balance sheet to deliver a best-in-class frontier model sans drama and throttling.

We released the portfolio ideas generation tool within Kilby. We also just released the below comparison function.

My Prompt to Kilby: “Compare the MD&A sections of the 2 attached 10K filings for Nvidia.”

Kilby’s response:

Here’s a direct comparison of the MD&A sections from both filings (FY2026 ended Jan 25, 2026 vs. FY2025 ended Jan 26, 2025):

Revenue Performance

FY2026 (from NVDA 10K 01252026.pdf, p. 37): Revenue hit $215.9B, up 65% YoY. Data Center up 68%, Gaming up 41%, Professional Visualization up 70%, Automotive up 39. Every segment accelerated.

FY2025 (from NVDA 10K 01262025.pdf, p. 38): Revenue was $130.5B, up 114% YoY. Data Center up 142%, driven by Hopper architecture for LLMs and generative AI. Gaming up only 9%. A much more concentrated growth story — Data Center did the heavy lifting while everything else was relatively flat.

The shift: FY2025 was a Data Center monoculture story. FY2026 broadened meaningfully — Gaming, ProViz, and Auto all put up strong numbers alongside continued Data Center dominance.

Architecture Transition

FY2025 MD&A: Hopper drove the full year; Blackwell only started shipping in Q4 FY2025.

FY2026 MD&A: Blackwell became the majority of Data Center revenue. Management highlighted the Blackwell transition as complex — product transitions and sophisticated system configurations introduced risk of delays, revenue volatility, inventory provisions, and warranty costs. They also flagged Rubin (next platform) with production shipments expected in H2 FY2027.

Gross Margin

FY2025: 75.0%, up 2.3 pts — driven by higher Data Center revenue mix.

FY2026: 71.1%, down 3.9 pts — the Hopper→Blackwell transition compressed margins as the company moved from selling HGX systems to full-scale datacenter solutions. A $4.5B H20 inventory charge hit the first quarter hard and was a primary drag.

The H20 / China Export Control Issue

This is a major new element in the FY2026 MD&A that was absent from FY2025. In April 2025, the USG required licenses for H20 exports to China and D:5 countries. NVIDIA took a $4.5B charge in Q1 FY2026. By the end of FY2026, management stated plainly that they were “effectively foreclosed from competing in China’s data center computing/compute market.” The FY2026 MD&A also flagged the Mellanox antitrust investigation in China, potential financial penalties, and the pending replacement rule for the AI Diffusion IFR.

FY2025 MD&A addressed export controls but with a different posture — the AI Diffusion IFR had just been published (Jan 15, 2025), its implications were forward-looking, and H20 was still being sold into China. The tone was cautionary but not yet describing a foreclosure.

Operating Expenses

FY2025: Opex $16.4B, up 45% — compensation, compute infrastructure, and new product engineering costs drove the increase.

FY2026: Opex $23.1B, up 41% — same drivers. R&D hit $18.5B (up 43%), with a 79% increase in compute and infrastructure costs specifically called out. SG&A up 31%.

Cash Flow and Liquidity

FY2025: Operating cash flow $64.1B; cash/securities $43.2B.

FY2026: Operating cash flow $102.7B — a significant jump driven by higher revenue. Cash/securities grew to $62.6B. Investing activity surged to ($52.2B), driven by $17.5B in private company/fund investments and the $13B Groq license deal. The Groq transaction (Dec 2025) — a non-exclusive license for LPU technology, $17B total consideration, $14.4B goodwill recorded — is a major new item with zero parallel in the FY2025 MD&A.

Capital Return

FY2025: $34.0B in buybacks; $834M dividends.

FY2026: $40.4B in buybacks; $974M dividends. Board approved a new $60B repurchase authorization in Aug 2025.

Structural Shifts in Tone and Risk Focus

The FY2025 MD&A was structured around three discrete challenge buckets: demand/supply complexity (Blackwell ramp), global trade (export controls as a forward risk), and macroeconomic factors. The narrative was optimistic — Blackwell was arriving, China was still a revenue contributor, and the AI Diffusion IFR was speculative.

The FY2026 MD&A is materially more sober in two areas: (1) China is now a lost market, with specific language about “effective foreclosure” and competitive harm to rivals benefiting from NVIDIA’s absence; and (2) the investment portfolio risk is substantially larger — $17.5B in illiquid private investments, $3.5B in facility lease guarantees, $11.4B in investment commitments, and a pending OpenAI deal. Management acknowledged these carry real downside if early-stage portfolio companies fail.

One Item to Flag

The FY2026 goodwill balance jumped from $5.2B to $20.8B — almost entirely due to the Groq license deal. That’s $14.4B of goodwill from a non-exclusive IP license where NVIDIA hired some employees but acquired no customer contracts or equity. The auditor flagged inventory valuation as the critical audit matter in both years, but the Groq structure — and the aggressive goodwill accounting — is worth watching. It uses a “cost to recreate” methodology which involves inherently unobservable inputs.

• Query “/ideas...” and you’ll be able to run a multi-variable query.

• The below metrics are trailing (TTM). I’m looking for a reasonably-priced source for forward estimates.

• If you have your own estimates, those can be plugged into Kilby.

• Any requests, let me know.

• I may push this test product out soon limited release. It will be a few weeks before we push the final product out.

example: /ideas technology companies with P/E above 0 and below 25 and EV/EBITDA above 0 and below 25: