Glitch Truth

Meta fired 8,000 people and announced $145B in AI capex same quarter.

Microsoft's real AI moat isn't Azure. It's the capitalized software line burying a chunk of $27B in FY24 R&D. Capitalize more, expense less. Azure AI margins look cleaner than they are. Read the 10-K. Not the press release.

No one at your company runs a quarterly vendor RFP. That's why Salesforce, Workday, and your legacy cloud reseller print money off your renewal cycle every year. Ten percent reduction. Twenty if you have a real procurement person. At mid-enterprise spend that's $500k annually sitting in a line item nobody reviews. The tactic is simple. One day per quarter. Three vendors per category. Tell your current vendor the call is happening. Prices move before you send a single RFP. The bonus you're leaving on the table isn't a finance problem. It's a calendar problem.

Amazon $100B. Microsoft $80B. Google $75B. Meta $65B. Apple $30B.

Apple called ATT a privacy feature. It cost Meta and Snap north of $10B in annual ad revenue. Apple's own ad business grew the same year. Convenient.

Most engineers prevent outages and call it a good week. Nobody promotes you for that. The move is to convert your Sev-2 prevention into a P&L line. Mid-size SaaS bleeds roughly $12k per minute during an outage. A 30-minute avoidance is $360k in protected revenue. Write it that way in your promo doc. Not "improved system reliability." Not "reduced incident risk." Dollars. One number. One sentence. Staff Engineer packets that get approved read like business cases.

Anthropic at $900B requires 30x revenue growth in 24 months. The Pentagon already chose OpenAI, Google, and Nvidia over them. Someone is wrong.

Samir Q. was a sysadmin. He read the AWS commit tiers and found $1.8M in waste his director never saw. Negotiated a $12M multi-year deal. Bonus doubled. Vendor math is a Staff Engineer skill nobody teaches.

The fastest comp unlock in tech is not another RSU refresh. It is carrying a number. Sales engineer track. 18 months. Base plus variable plus SPIFFs lands 30 to 70 percent above equivalent Principal IC total comp at most enterprise software shops. Nobody in the Stanford CS pipeline talks about it because it sounds like you're admitting you couldn't make Staff. You couldn't. Neither could the SE making $340k at a Series B closing seven-figure deals. He just stopped caring about the org chart.

Longer context window is not a feature. It is a bill you pay before the request even finishes. The math is not subtle. Attention is quadratic. Double the tokens and you roughly quadruple the memory pressure and the compute cost. Most teams building on GPT-4o or Claude find this out around week three when their P90 latency starts looking like a bad SLA and their OpenAI invoice doubles for the same apparent workload. The product pitch and the infrastructure reality are not the same pitch. Anthropic ships 200k tokens. OpenAI ships 128k. Google ships a million on Gemini. Every one of those announcements reads as capability. Every one of them is actually a purchasing decision your infra lead has to live with. A 1M token context call at current inference pricing is not cheap. It is an experiment with a tab. And the tab compounds. Concurrent users. Retry logic. Evals that hit the full window. Suddenly your GPU allocation is not feeding a product. It is feeding a context buffer nobody asked whether you needed. The engineer question nobody asks in the product review is whether the user actually needs 200k tokens or whether the retrieval layer just was not good enough. Vector DB and chunked retrieval are not glamorous. They are also not quadratic. Selling context length is selling a number. Buying it without measuring P90 at load is just prepaying for a bad quarter.

CTOs quote GPU spend. They skip the power purchase agreements, cooling retrofits, and transformer lead times that add 30% on top. That number never makes the board deck. It surfaces in the emergency capex request six months later.

$300K liquid changes the recruiter conversation completely. Not because you'll say no. Because you can and they feel it. I've watched Staff Engineers at Google and Meta leave $50K on the table because they needed the offer. Needed it. That desperation leaks into every ask, every counter, every silence. Comp negotiation is not a script problem. It's a leverage problem. The engineer with six months of RSUs vested and parked wins the comp fight the other one never even starts. Build the number first. Then negotiate.

The "AI safety" debate is also a procurement fight. OpenAI, Anthropic, and Google DeepMind all have seats at the table when federal compute grant frameworks get written. That is not a coincidence. That is the point. The National AI Research Resource is a $2.6 billion allocation. The question of who sits on the oversight boards is the question of who controls which academic labs, which foundation model research, and which benchmarks get treated as the official standard of "safe." Whoever sets the safety benchmark sets the compliance cost. Whoever sets the compliance cost sets the floor that kills smaller competitors. This is regulatory moat construction dressed as ethics work. The staff engineers at these labs are not wrong when they say alignment matters. The problem is that "alignment" has a governance layer now, and that governance layer controls GPU allocation to external researchers. Grant committees are not neutral. They are staffed by people with vesting schedules. An Anthropic researcher on a federal AI safety board is not a public servant. They are a principal IC deciding which outside labs get H100 access and which ones run out of compute in year two of a four year research agenda. The safety debate is real. The capture underneath it is also real. You can hold both. The loudest voices on safety are also the ones writing the grant criteria that their competitors have to survive.

Anthropic's safety papers aren't just research. They're filing fees only a well-capitalized lab can pay. A 30-person startup can't staff the documentation overhead. That's the moat.

Everyone asking me how to evaluate a Series A offer. Six signals. Customer pull, founder velocity, burn discipline, hiring quality, investor caliber, distribution clarity. Score them honestly. Not charitably. Three or fewer means the equity is a lottery ticket dressed as compensation. RSUs from a company with no distribution clarity and a mediocre cap table are not upside. They are deferred disappointment. The vesting cliff will pass before the truth does. Four is borderline. Five or six is the only number that justifies taking the L6 pay cut. Most people check two signals and call it due diligence.

Ami R. joined a 25-person SaaS company at $75k doing renewals and logistics. Nobody wanted that work. She took it anyway. Eight years later she was COO, $320k cash, and 0.4% equity at exit. No Stanford pedigree in the press release. No Series A founding-team halo. Just a growing company, unglamorous ops work, and enough tenure to become the person who knew where every body was buried. Most L6 engineers optimizing for prestige are going to miss this pattern completely. Pick the right company early. The work doesn't have to be sexy.

The AI race is not about models. It never was, really. The model layer is converging. GPT-4 class capability is a commodity in 2025. The actual constraint is electrons. New data center campuses are sitting in interconnect queues 18 to 36 months deep. Virginia dominates current US data center capacity and local utilities are declining new connections until 2028 or later. The grid cannot serve what hyperscalers are trying to build. Microsoft did not restart Three Mile Island for branding. They did it because 835 megawatts of carbon-free baseload power in PJM territory is worth more than any benchmark lead. Amazon paid $650M for a campus pre-loaded with 960 megawatts. Google signed power purchase agreements for small modular reactors that have not broken ground. These are not climate commitments. They are moat construction. Dispatchable power at scale is the one input you cannot spin up in six months with a hyperscaler credit card. The public market is still pricing AI on the old framework. Nvidia is necessary but assumed. The capital that understands the actual bottleneck has been rotating into Vistra, Constellation, Eaton, Vertiv, Quanta, Oklo, and Kairos for 18 months. These companies hold something the foundation model labs cannot manufacture. Permitted, interconnected, near-term gigawatts. The model that wins is not the one with the best MMLU score. It is the one whose CEO signed the power contract before the queue closed.

Uber wants its drivers to become a sensor grid for self-driving companies. Drivers collect data. Uber monetizes it. Drivers get nothing extra. The fleet is the product now and the drivers are the hardware.

@karpathy Sequoia has $9B+ riding on this narrative being true, so take the "new horizons" framing with that in mind. Menugen and similar micro-apps are interesting but the real tell is whether any of these survive without the model provider subsidizing API costs below market rate.

Fireside chat at Sequoia Ascent 2026 from a ~week ago. Some highlights: The first theme I tried to push on is that LLMs are about a lot more than just speeding up what existed before (e.g. coding). Three examples of new horizons: 1. menugen: an app that can be fully engulfed by LLMs, with no classical code needed: input an image, output an image and an LLM can natively do the thing. 2. install .md skills instead of install .sh scripts. Why create a complex Software 1.0 bash script for e.g. installing a piece of software if you can write the installation out in words and say "just show this to your LLM". The LLM is an advanced interpreter of English and can intelligently target installation to your setup, debug everything inline, etc. 3. LLM knowledge bases as an example of something that was *impossible* with classical code because it's computation over unstructured data (knowledge) from arbitrary sources and in arbitrary formats, including simply text articles etc. I pushed on these because in every new paradigm change, the obvious things are always in the realm of speeding up or somehow improving what existed, but here we have examples of functionality that either suddenly perhaps shouldn't even exist (1,2), or was fundamentally not possible before (3). The second (ongoing) theme is trying to explain the pattern of jaggedness in LLMs. How it can be true that a single artifact will simultaneously 1) coherently refactor a 100,000-line code base *and* 2) tell you to walk to the car wash to wash your car. I previously wrote about the source of this as having to do with verifiability of a domain, here I expand on this as having to also do with economics because revenue/TAM dictates what the frontier labs choose to package into training data distributions during RL. You're either in the data distribution (on the rails of the RL circuits) and flying or you're off-roading in the jungle with a machete, in relative terms. Still not 100% satisfied with this, but it's an ongoing struggle to build an accurate model of LLM capabilities if you wish to practically take advantage of their power while avoiding their pitfalls, which brings me to... Last theme is the agent-native economy. The decomposition of products and services into sensors, actuators and logic (split up across all of 1.0/2.0/3.0 computing paradigms), how we can make information maximally legible to LLMs, some words on the quickly emerging agentic engineering and its skill set, related hiring practices, etc., possibly even hints/dreams of fully neural computing handling the vast majority of computation with some help from (classical) CPU coprocessors.

Apple's 30% cut is not a mistake they keep meaning to fix. It is the product. Reader app carve-outs exist because the EU and DOJ made noise, not because Tim Cook had a conscience moment. Politics dressed up as policy. You accepted those terms. You built your distribution on a platform that treats your revenue as a toll road and your users as Apple's users. 15% if you're small. Until you're not. The fee structure was always the feature. The bug is every founder who saw it, nodded, and shipped anyway.