2025-07-10 08:00:00
For the last decade, the biggest line item in any startup’s R&D budget was predictable talent. But AI is pushing its way onto the P&L.
How much should a startup spend on AI as a percentage of its research and development spend?
10%? 30%? 60?
There are three factors to consider. First, the average salary for a software engineer in Silicon Valley. Second is the total cost of AI used by that engineer. Cursor is now at $200 per month for their Ultra Plan & reviews of Devin suggest $500 per month. Third, the number of agents an engineer can manage.
A first pass :
| Item1 | Cost/Year |
|---|---|
| Software Engineer Salary | $200,000 |
| Agent Subscription | $18,000 ($6,000 x 3) |
| Total Cost | $218,000 |
| AI % of Total | 9% |
But the subscription costs are probably low. Over the last few days I’ve been playing around extensively with AI coding agents and I racked up a bill of $1,000 within the span of five days! 😳😅
So let’s update the table and assume another $1000 per month per engineer.
| Item | Cost/Year |
|---|---|
| Software Engineer Salary | $200,000 |
| Agent Subscription | $18,000 |
| Additional AI Costs2 | $12,000 |
| Total Cost | $230,000 |
| AI % of Total | 13% |
So for a typical startup, an estimate of 10 to 15% of total R&D expense today might conceivably be used for AI.
The variants will be much broader in practice as we all learn to use AI better and it penetrates more of the organization. Smaller companies that are AI native from the outset are likely to have significantly higher ratios.
If you’re interested to participate in an anonymous survey, I will be publishing the results if the sample size is sufficiently large to have a statistically significant result.
2025-07-08 08:00:00
When you query AI, it gathers relevant information to answer you.
But, how much information does the model need?
Conversations with practitioners revealed the their intuition : the input was ~20x larger than the output.
But my experiments with Gemini tool command line interface, which outputs detailed token statistics, revealed its much higher.
300x on average & up to 4000x.
Here’s why this high input-to-output ratio matters for anyone building with AI:
Cost Management is All About the Input. With API calls priced per token, a 300:1 ratio means costs are dictated by the context, not the answer. This pricing dynamic holds true across all major models.
On OpenAI’s pricing page, output tokens for GPT-4.1 are 4x as expensive as input tokens. But when the input is 300x more voluminous, the input costs are still 98% of the total bill.
Latency is a Function of Context Size. An important factor determining how long a user waits for an answer is the time it takes the model to process the input.
It Redefines the Engineering Challenge. This observation proves that the core challenge of building with LLMs isn’t just prompting. It’s context engineering.
The critical task is building efficient data retrieval & context - crafting pipelines that can find the best information and distilling it into the smallest possible token footprint.
Caching Becomes Mission-Critical. If 99% of tokens are in the input, building a robust caching layer for frequently retrieved documents or common query contexts moves from a “nice-to-have” to a core architectural requirement for building a cost-effective & scalable product.
For developers, this means focusing on input optimization is a critical lever for controlling costs, reducing latency, and ultimately, building a successful AI-powered product.
2025-07-05 08:00:00
If I have a dollar to invest in a stock or a crypto token, how do I decide? I need to compare across the two.
Historically, that comparison was impossible. Crypto traded on a potent cocktail of hype, narrative, & the promise of a decentralized future. Perception drove valuations.
That’s changing. The word “revenue” is no longer verboten in the world of crypto. It’s becoming the goal.
This trend will unlock the next wave of institutional capital because investors can compare the risk/reward of crypto with the same metrics as other software companies.
Look at Hyperliquid, a decentralized options/perpetuals exchange. It’s on a $650M+ annualized revenue run rate, trading at a 60x multiple. This valuation is rich compared to public fintech companies like Coinbase (13x) or Robinhood (24x), but it’s based on comparable tangible financial performance.
The same is true for blockchains: Optimism & Arbitrum trade at 40-60x revenues. Amazon Web Services or Azure, if they were to be spun out as separate entities, would trade at 20-30x.
Phantom, a brokerage app1, has generated $394m in lifetime revenue, most of it in the last 6 months.
The list is still small. Fewer than 40 apps generate more than $1m. But the ones that do can generate hundreds of millions.
This is the future of crypto. A future where revenue generation is the goal.
This does not mean the attention economy is dead. Hype will always be part of crypto. But the projects that attract the next wave of capital will combine a compelling narrative with a sustainable business model. They will show a clear path to revenue & trade on revenue multiples - likely elevated relative to the rest of software because of the explosive growth potential.
The message for builders is clear: focus on fundamentals. Build a product people want, that solves a real problem, & that generates real revenue. That revenue provides the capital to fuel crypto’s next chapter.
Read the original post that inspired this one here.
I know it’s called a wallet, but I suspect wallets will be renamed brokerages as Robinhood & Coinbase fuse stocks & crypto. ↩︎
2025-07-02 08:00:00
Yesterday, Figma filed its beautifully designed S-1.
It reveals a product-led growth (PLG) business with a remarkable trajectory. Figma’s collaborative design tool platform disrupted the design market long-dominated by Adobe.
Here’s how the two companies stack up on key metrics for their most recent fiscal year:
| Metric (2024) | Figma | Adobe |
|---|---|---|
| Revenue (YoY Growth) | $749M (48%) | $21.5B (11%) |
| Gross Margin | 88.3% | 89.0% |
| Non-GAAP Op Margin | 17.0% | 44.5% |
| Sales Efficiency | 1.00 | 0.39 |
| Adjusted FCF Margin1 | 24.2% | 36.6% |
| Net Dollar Retention | 132% | NA |
| Customers > $100k ARR | 963 | NA |
Figma is about 3% the size of Adobe but growing 4x faster. The gross margins are identical. Figma’s 132% Net Dollar Retention is top decile.
The data also shows Figma’s Research & Development spend nearly equals Sales & Marketing spend.
This is the PLG model at its best. Figma’s product is its primary marketing engine. Its collaborative nature fosters viral, bottoms-up adoption, leading to a best-in-class sales efficiency of 1.0. For every dollar spent on sales & marketing in 2023, Figma generated a dollar of new gross profit in 2024. Adobe’s blended bottoms-up & sales-led model yields a more typical 0.39.
The S-1 also highlights risks. The most significant is competition from AI products. While Figma is investing heavily in AI, the technology lowers the barrier for new entrants. Figma’s defense is its expanding platform—with products like FigJam, Dev Mode, & now Slides, Sites, & Make.
These new product categories have driven many PLG AI software companies to tens & hundreds of millions in ARR in record time.
Given its high growth & unique business model, how should the market value Figma? We can use a linear regression based on public SaaS companies to predict its forward revenue multiple. The model shows a modest correlation between revenue growth & valuation multiples (R² = 0.23).
Figma, with its 48% growth, would be the fastest-growing software company in this cohort setting aside NVIDIA. A compelling case can be made that Figma should command a higher-than-predicted valuation. Its combination of hyper-growth, best-in-class sales efficiency, & a passionate, self-propagating user base is rare.
Applying our model’s predicted 19.9x multiple to estimate forward revenue yields an estimated IPO valuation of approximately $21B2 - a premium to the $20B Adobe offered for the company in 2022.
The S-1 tells the story of a category-defining company that built a collaborative design product, developed a phenomenal PLG motion, & is pushing actively into AI.
The $1.0 billion termination fee from Adobe was received in December 2023 and recorded as “Other income, net” in Fiscal Year 2024 (ending January 31, 2024). The large stock-based compensation charge of nearly $900 million is related to an employee tender offer in May 2024. Both of these are removed in the non-GAAP data cited above. ↩︎
By taking Figma’s 48.3% trailing twelve-month growth rate & discounting it by 15% (to account for a natural growth slowdown), the model produces a forward growth estimate of 41.1%. This would imply forward revenue of about $1.1b. ↩︎
2025-06-27 08:00:00
Gmail’s AI email assistant writes like a committee of lawyers designed it.
Pete Koomen’s recent post Horseless Carriages explains why: developers control the AI prompts instead of users. In his post he argues that software developers should expose the prompts and the user should be able to control it.
He inspired me to build my own. I want a system that’s fast, accounts for historical context, & runs locally (because I don’t want my emails to be sent to other servers), & accepts guidance from a locally running voice model.1
Here’s how it works,
Here are two examples : emailing a colleague, Andy, & a hypothetical founder.
Instead of generics, the system learns from my actual email history. It knows how I write to investors vs colleagues vs founders because it’s seen thousands of examples.
The point isn’t that everyone will build their own email system. It’s that these principles will reshape software design.
We’re still in the horseless carriage era of AI applications.
The breakthrough will come when software adapts to us instead of forcing us to adapt to it.
1 Centered around a command line email client called Neomutt.
2 The software hits LanceDB, a vector database with embedded emails & finds the ones that are the most relevant from the sender to match the tone.
3 The code is here.
2025-06-25 08:00:00
In “Data Rules Everything Around Me,” Matt Slotnick wrote about the difference between SaaS & AI apps. A typical SaaS app has a workflow layer, a middleware/connectivity layer, & a data layer/database. So does an AI app.
AI makes writing frontends trivial, so in the three-layer cake of workflow software the data matters much more.
The big differences between an AI & the SaaS app lie within the ganache of the middle layer. In SaaS applications, coded business rules determine each step a lead follows from creation to close.
In AI apps, a non-deterministic AI model decides the steps using context : relevant information about the lead that the AI is querying from other sources.
The better the data, the better the workflow.
The context is the most valuable component because it ultimately changes the workflow. Models are relatively similar in performance.
For example, an inbound email comes into a customer support desk, “Was I double charged this month?” An agentic workflow would query the billing system, the contract system, & the email drafting tool to draft an email to the customer with distinct language for that persona. This only works if the enterprises’ data is well structured.
Enterprises will be shy about sharing the context with their vendors because of how much value it provides. They may start to structure it & assign a department to manage it because the better its availability, the more effective the agentic systems will be.
Data architecture may become a competitive advantage & the future battleground for software companies will be the access to that context - & the fight has already begun.
