Tomasz TunguzModify

2025-10-07 08:00:00

Chris Klayko brings over two decades of sales leadership experience transforming technology companies from promising startups to multi-billion dollar enterprises. As SVP of Sales at Databricks, Chris leads the charge in democratizing data & AI for organizations worldwide. His remarkable journey includes scaling Google Cloud from tens of millions to a multi-billion dollar business in just four years, driving UiPath’s Americas expansion during its hypergrowth phase, & building SAP’s emerging solutions division.

On October 15th in San Francisco, Theory Ventures is hosting an exclusive Office Hours session with Chris Klayko, SVP of Sales at Databricks.

During this intimate fireside chat, Chris & I will explore :

• Building & motivating high-performing sales teams in the AI era
• Lessons from scaling organizations from $100M → $1B+ at Databricks, UiPath & Google Cloud
• How sales leaders can adapt in a fast-changing, AI-driven landscape
• The evolution from traditional software sales to AI & data platform selling
• Creating sales cultures that attract & retain top talent
• Navigating the shift from product-led to enterprise sales motions

San Francisco | October 15, 5:30–7:30pm | Space is limited

Apply to join here. Submit your questions through the registration form & I’ll weave them into our conversation.

I look forward to welcoming Chris to Office Hours & sharing his insights with our community!

2025-10-06 08:00:00

Last Saturday I spent three hours implementing a coding workflow I’d read about that morning. By Sunday afternoon I’d tested it on two projects , measured the results , & decided to keep it.

Ten years ago , the same journey would have taken weeks. Buy a Getting Things Done book , set up a Bullet Journal , or learn Zettelkasten. Read it. Understand the methodology. Build the habits. Iterate through mistakes. Maybe after a month you’d know if it worked.

Something fundamental shifted. We moved from commitment-based learning to experiment-based discovery. Three recent experiments show what’s possible.

First, Paper2Agent transforms research papers into working code agents in 30 minutes to 3 hours. Ask it to implement AlphaGenome’s genomic analysis from a published paper & it processes the code, builds the environment, runs the tutorials. Previously: read the paper (2 days), understand the methodology (1 day), code the implementation (3 days). Now: point AI at the repository, get a working agent by lunch.

Second, when I discovered Google’s Agent Development Kit research on tool design patterns, I asked AI to redesign my automation tools following those principles. The result : a 41% reduction in AI operation costs, implemented while I was answering emails in another browser tab.

Third, over the weekend I read Jesse Vincent’s architect/implementer workflow via Simon Willison’s write-up. The approach splits AI coding into two sessions: one architect to design, one implementer to execute. No weeks of trial & error to understand the nuances. Read, implement, test, done.

The pattern emerging : instant implementation.

We can now test productivity systems like trying coffee drinks. Monday : architect/implementer (cappuccino—structured , classic). Tuesday : traditional single-session (flat white—simpler , maybe better?). Wednesday : hybrid workflow (chai latte—wait , is this even coffee?). By Thursday , the data shows which saves 3 hours per week.

The self-help aisle promised transformation through commitment. AI promises transformation through experimentation.

2025-10-03 08:00:00

When Nvidia announced a $100 billion investment commitment to OpenAI¹ in September 2025 , analysts immediately drew comparisons to the telecom bubble. The concern : is this vendor financing , where a supplier lends money to customers so they can buy the supplier’s products , a harbinger of another spectacular collapse?

American tech companies will spend $300-400 billion on AI infrastructure in 2025^2,3 , exceeding any prior single-year corporate infrastructure investment in nominal dollars.³ David Cahn estimates the revenue gap has grown to $600 billion⁴.

I analyzed the numbers. The similarities are striking , but the differences matter.

The Lucent Playbook

In 1999 , Lucent Technologies reached $37.92 billion in revenue at the peak of the dot-com bubble. ⁵ Lucent was the #1 North American telecommunications equipment manufacturer with 157,000 employees & dominated markets alongside Nortel Networks (combined 53% optical transport market share). ⁶ Behind the scenes , equipment makers extended billions in vendor financing to telecom customers. Lucent committed $8.1B⁷ , Nortel extended $3.1B with $1.4B outstanding , & Cisco promised $2.4B in customer loans.⁸

The strategy seemed brilliant : lend money to cash-strapped telecom companies so they could buy your equipment. Everyone wins—until the merry-go-round stops.

When the bubble burst :

• 47 Competitive Local Exchange Carriers (CLECs) bankrupted 2000-2003 , including Covad , Focal Communications , McLeod , Northpoint , Winstar ^9,10
  • Why they failed : $60B overbuild 1996-2001 , market saturation from identical business models , sudden funding collapse (Jan 2001 : billions available , Apr 2001 : zero)¹¹
• 33-80% of vendor loan portfolios went uncollected as customers failed & equipment became worthless¹²
• Fiber networks were using less than 0.002% of available capacity , with potential for 60,000x speed increases. ¹³ It was just too early.

Nvidia’s Playbook

Fast forward to 2025. Nvidia’s vendor financing strategy totals $110 billion in direct investments plus another $15+ billion in GPU-backed debt. The largest commitment is $100B to OpenAI (September 2025)^1,14 , structured as 10 tranches of $10B each tied to infrastructure deployment milestones. The first $10B was valued at a $500B OpenAI valuation , with subsequent tranches priced at prevailing valuations. Payment comes via lease arrangements , not upfront GPU purchases. OpenAI CFO Sarah Friar confirmed : “Most of the money will go back to Nvidia”¹⁴

Beyond OpenAI , Nvidia holds a $3B stake in CoreWeave¹⁵ , a company that has spent $7.5B on Nvidia GPUs , & $3.7B in other AI startup investments¹⁶ through NVentures.

The GPU-backed debt market adds another layer. CoreWeave alone carries $10.45B in debt using GPUs as collateral¹⁷. An additional $10B+ in GPU-backed debt has emerged for “Neoclouds” including Lambda Labs ($500M GPU-backed loan)^18,19.

Lucent in 1999-2000 had vendor financing commitments of $8.1B (24% of $33.6B revenue). Nvidia’s direct investments total 67% of annual revenue ($110B against $165B LTM). Nvidia’s exposure is 2.8x larger relative to revenue than Lucent’s official outstanding loans , though Lucent’s off-balance-sheet guarantees masked the true exposure.

The Numbers Side-by-Side (2024 Dollars)

The Reasons to be Wary

1. The AI Customer Base is More Concentrated

Lucent’s top 2 customers—AT&T at 10% & Verizon at 13%—accounted for 23% of revenue in FY2000.²¹ The Regional Bell Operating Companies , or RBOCs , the seven “Baby Bells” created from AT&T’s 1984 breakup , were also major customers. Nvidia has 39% of revenue from just 2 customers & 46% from 4 customers , nearly double Lucent’s concentration. 88% of Nvidia’s revenue comes from data centers.

2. GPU-Backed Debt Is New

The new $10B+ GPU-backed debt market is built on the assumption that GPUs will hold their value over 4-6 years. GPU-backed loans carry ~14% interest rates²² , triple investment-grade corporate debt.²³

How Depreciation Schedules Changed :

Amazon’s 2025 reversal (6 → 5 years) is the first major pullback.

CPUs historically have 5-10 years of useful life , while GPUs in AI datacenters last 1-3 years in practice , despite 6-year accounting assumptions.^30,31 Evidence from Google architects shows GPUs at 60-70% utilization survive 1-2 years , with 3 years maximum.³¹ Meta’s Llama 3 training experienced 9% annual GPU failure rates , suggesting 27% failure over 3 years.³¹

Cerno Capital raises the question : “Are these policies a reflection of genuine economic & technological realities? Or are these policies a lever by which hyperscalers are enhancing the optics of their investment programs amid rising investor concerns?”³²

4. The Use of SPVs

Tech companies use Special Purpose Vehicles (SPVs) to finance AI datacenter construction. A hyperscaler like Meta partners with a private equity firm like Apollo , contributing capital to a separate legal entity that builds & owns the datacenter.

As investor Paul Kedrosky explains : “I have a stake in it as Meta. Some giant private debt provider has a stake in it. The datacenter is under my control. But I don’t own it, so you don’t get to roll it back into my balance sheet.”²*

The Structure

1. Entity Creation : Hyperscaler & PE firm form separate legal entity (SPV)
2. Capital Structure : Typically 10-30% equity, 70-90% debt from private credit markets
3. Lease Agreement : SPV leases capacity back to hyperscaler
4. Balance Sheet Treatment : SPV debt doesn’t appear on hyperscaler’s balance sheet

The hyperscaler maintains operational control through long-term lease agreements. Because it doesn’t directly own the SPV , the debt remains off its balance sheet under current accounting standards.

The appeal is straightforward. “I don’t want the credit rating agencies to look at what I’m spending. I don’t want investors to roll it up into my income statement.”²*

Market Scale

American tech companies are projected to spend $300-400 billion on AI infrastructure in 2025. Hyperscaler capital expenditures have reached approximately 50% of operating income², levels historically associated with government infrastructure buildouts rather than technology companies.

Where the Risk Sits

Datacenter assets now represent 10-22% of major REIT portfolios² , up from near zero two years ago. The thin equity layer (10-30%) means if datacenter utilization falls short of projections or if GPUs depreciate faster than projected , equity holders face losses before debt holders experience impairment.

*Quotes lightly edited for clarity & brevity

5. Custom Silicon Threat

Hyperscalers are building their own AI accelerators to reduce Nvidia dependence. Microsoft aims to use “mainly Microsoft silicon” , specifically Maia accelerators , in datacenters.³³ Google deploys TPUs , Amazon builds Trainium & Inferentia chips , & Meta develops MTIA processors. If customers shift to in-house silicon , CoreWeave’s GPU collateral value & Nvidia’s vendor financing become exposure to customers building competitive alternatives.

Nvidia Isn’t Lucent & 2025 Isn’t 2000

• Accounting : Lucent manipulated $1.148B in revenue , SEC charged 10 executives with fraud⁵ ; Nvidia shows no evidence of manipulation , audited by PwC , Aa3 rated³⁴
• Cash flow : Lucent lent $8.1B while cash flow lagged profitability & receivables exploded $5.4B (1998-1999)²⁰ ; Nvidia lends with $50B+ annual operating cash flow & $46.2B net cash³⁵
• Credit rating : Lucent downgraded to A3 (December 2000)³⁶ ; Nvidia upgraded to Aa3 (March 2024)³⁴
• Customer base : Lucent’s customers were leveraged CLECs burning capital ; Nvidia’s top 4 customers generated $451B in operating cash flow in 2024 (Microsoft $119B , Alphabet $125B , Amazon $116B , Meta $91.3B)³⁷
• Capacity : Fiber networks used <0.002% of capacity in 2000¹³ ; Microsoft & AWS report AI capacity constraints in 2025^38,39

What I’m Watching

Is AI demand real (like cloud computing) or speculative (like dot-com fiber)?

Here’s what I’m watching :

1. GPU utilization rates : Are data centers actually using the chips or just stockpiling?
2. OpenAI’s monetization : Can they generate enough revenue to justify the buildout?
3. Debt defaults : Any cracks in the $15B GPU-backed debt market?
4. AR trends : AR improved from 68% (FY24) to 30% (Q2 FY26) , but still watch for deterioration
5. Customer adds : Are new customers emerging , or is Nvidia dependent on the same 2-4 hyperscalers?
6. Custom silicon threat : Microsoft developing Maia accelerators , aiming to use “mainly Microsoft silicon in the data center.”³³ If hyperscalers shift to in-house chips , Nvidia’s vendor financing becomes exposure to customers building competitive alternatives.
7. Vendor consolidation : Many companies are in a period of experimentation trying 2-3 competing vendors. Those experimental budgets may thin with time , reducing overall spend.

AI is already broadly deployed—40% of US employees used AI at work by September 2025 , double the 20% rate in 2023.⁴⁰ Questions persist about effectiveness : the oft cited MIT study found 95% of AI pilots failed to deliver measurable P&L impact , primarily due to poor integration rather than technical failures.⁴¹

Yet the pace of improvement is tremendous. Labor market data shows wages rising twice as fast in AI-exposed industries , & workers using AI boost performance up to 40%.⁴⁰ Many of Nvidia’s customers are profitable & sophisticated hyperscalers—Microsoft , Google , Amazon , Meta—generating $451B in operating cash flow in 2024 , with tremendous pull from their own enterprise customers demanding AI. OpenAI is not profitable , reporting a $4.7B loss in H1 2025 on $4.3B revenue , though nearly half the loss is stock-based compensation.⁴²

Unlike the telecom bubble , where demand was speculative & customers burned cash , this merry-go-round has paying riders.

Coda : Lucent’s Accounting Fraud

Behind the vendor financing disaster was systematic accounting fraud. The SEC charged Lucent with manipulating $1.148 billion in revenue & $470 million in pre-tax income during fiscal year 2000. ⁵ The fraud involved multiple schemes :

Channel Stuffing : Lucent sent $452 million in equipment to distributors but counted it as revenue before the distributors sold to end customers.⁵ This created phantom sales.

Side Agreements : Lucent executives entered secret agreements with distributors granting them return rights & privileges beyond their distribution contracts , making it improper to recognize revenue.⁵ These side deals were hidden from auditors.

Reserve Manipulation : Lucent improperly established & maintained excess reserves to smooth earnings , violating GAAP.⁵

The SEC charged 10 Lucent executives with securities fraud.⁵ The company paid a $25 million fine—the largest ever for failing to cooperate with an SEC investigation.⁵ The accounting manipulation masked deteriorating fundamentals until too late.

The WinStar Collapse : Lucent committed $2 billion in vendor financing to WinStar Communications , a CLEC. When WinStar struggled , Lucent refused a final $90 million loan extension. WinStar filed bankruptcy. Lucent wrote off $700 million in bad debts.⁴³ This pattern repeated across customer defaults : Lucent made provisions for bad debts of $2.2 billion (2001) & $1.3 billion (2002)—a total of $3.5 billion in customer loan losses.⁴³

References

2025-10-02 08:00:00

AI breaks the data stack.

Most enterprises spent the past decade building sophisticated data stacks. ETL pipelines move data into warehouses. Transformation layers clean data for analytics. BI tools surface insights to users.

This architecture worked for traditional analytics.

But AI demands something different. It needs continuous feedback loops. It requires real-time embeddings & context retrieval.

Consider a customer at an ATM withdrawing pocket money. The AI agent on their mobile app needs to know about that $40 transaction within seconds. Data accuracy & speed aren’t optional.

Netflix rebuilt their entire recommendation infrastructure to support real-time model updates¹. Stripe created unified pipelines where payment data flows into fraud models within milliseconds².

The modern AI stack requires a fundamentally different architecture. Data flows from diverse systems into vector databases, where embeddings & high-dimensional data live alongside traditional structured data. Context databases store the institutional knowledge that informs AI decisions.

AI systems consume this data, then enter experimentation loops. GEPA & DSPy enable evolutionary optimization across multiple quality dimensions. Evaluations measure performance. Reinforcement learning trains agents to navigate complex enterprise environments.

Underpinning everything is an observability layer. The entire system needs accurate data & fast. That’s why data observability will also fuse with AI observability to provide data engineers & AI engineers end-to-end understanding of the health of their pipelines.

Data & AI infrastructure aren’t converging. They’ve already fused.

References

2025-09-30 08:00:00

I discovered I was designing my AI tools backwards.

Here’s an example. This was my newsletter processing chain : reading emails, calling a newsletter processor, extracting companies, & then adding them to the CRM. This involved four different steps, costing $3.69 for every thousand newsletters processed.

Before: Newsletter Processing Chain

# Step 1: Find newsletters (separate tool)
ruby read_email.rb --from "[email protected]" --limit 5
# Output: 340 tokens of detailed email data

# Step 2: Process each newsletter (separate tool)
ruby enhanced_newsletter_processor.rb
# Output: 420 tokens per newsletter summary

# Step 3: Extract companies (separate tool)
ruby enhanced_company_extractor.rb --input newsletter_summary.txt
# Output: 280 tokens of company data

# Step 4: Add to CRM (separate tool)
ruby validate_and_add_company.rb startup.com
# Output: 190 tokens of validation results

# Total: 1,230 tokens, 4 separate tool calls, no safety checks
# Cost: $3.69 per 1,000 newsletter processing workflows

Then I created a unified newsletter tool which combined everything using the Google Agent Development Kit, Google’s framework for building production grade AI agent tools :

# Single consolidated operation
ruby unified_newsletter_tool.rb --action process \
  --source "techcrunch" --format concise \
  --auto-extract-companies
# Output: 85 tokens with all operations completed

# 93% token reduction, built-in safety, cached results
# Cost: $0.26 per 1,000 newsletter processing workflows
# Savings: $3.43 per 1,000 workflows (93% cost reduction)

Why is the unified newsletter tool more complicated?

It includes multiple actions in a single interface (process, search, extract, validate), implements state management that tracks usage patterns & caches results, has rate limiting built in, & produces structured JSON outputs with metadata instead of plain text.

But here’s the counterintuitive part : despite being more complex internally, the unified tool is simpler for the LLM to use because it provides consistent, structured outputs that are easier to parse, even though those outputs are longer.

To understand the impact, we ran tests of 30 iterations per test scenario. The results show the impact of the new architecture :

We were able to reduce tokens by 41% (p=0.01, statistically significant), which translated linearly into cost savings. The success rate improved by 8% (p=0.03), & we were able to hit the cache 30% of the time, which is another cost savings.

While individual tools produced shorter, “cleaner” responses, they forced the LLM to work harder parsing inconsistent formats. Structured, comprehensive outputs from unified tools enabled more efficient LLM processing, despite being longer.

My workflow relied on dozens of specialized Ruby tools for email, research, & task management. Each tool had its own interface, error handling, & output format. By rolling them up into meta tools, the ultimate performance is better, & there’s tremendous cost savings. You can find the complete architecture on GitHub.

2025-09-29 08:00:00

“The way to do a piece of writing is three or four times over, never once.”

Writing is hard. John McPhee, who invented literary nonfiction that reads like a novel, developed a four-draft writing method that transforms chaotic ideas into compelling narratives.

McPhee pioneered creative nonfiction at The New Yorker, writing books like Oranges & Coming into the Country that made complex subjects fascinating through storytelling. His approach differs from traditional journalism by incorporating fiction techniques while maintaining factual accuracy. His prose combines vivid imagery with economy :

“The doctor listens in with a stethoscope and hears sounds of a warpath Indian drum.”

He favored directness :

“He liked to go from A to B without inventing letters between.”

About his genre, McPhee said :

“Nonfiction—what the hell, that just says, this is nongrapefruit we’re having this morning.”

McPhee later codified his approach in Draft No. 4: On the Writing Process, sharing decades of writing wisdom.

His organizational philosophy shapes everything :

“You can build a structure in such a way that it causes people to want to keep turning pages. A compelling structure in nonfiction can have an attracting effect analogous to a story line in fiction. Readers are not supposed to notice the structure. It is meant to be about as visible as someone’s bones.”

McPhee’s Four-Draft Framework :

1. Brain dump draft - Capture every possible idea, fact, & angle without editing or judgment
2. Structure draft - Organize ideas into logical sequences & identify the core narrative thread
3. Ruthless cut draft - Remove everything that doesn’t serve the primary message or confuse the reader
4. Polish draft - Refine prose, fix grammar, & ensure each sentence drives toward your goal

This is one of the best techniques I’ve found for writing. The method works because it separates creative thinking from critical evaluation. When you try to write perfect prose while generating ideas, it’s easy to fall into creative block.

Each draft becomes the foundation for the next, creating a recursive process that transforms chaotic thoughts into structured narratives. Like peeling back the layers of an orange to reveal the fruit within, each draft strips away what doesn’t belong, revealing the essential story that was always there waiting to be discovered.