ExpertsFMAI Enterprise - Databricks & Glean | BG2 Guest Interview
BG2Pod with Brad Gerstner and Bill Gurley · December 23, 2025
5 Selected Lessons:
framework · 2 heuristic · 1 priority · 1 prediction · 1
Value accrual ranking across data, intelligence, and application layers
Source
BG2Pod with Brad Gerstner and Bill Gurley — AI Enterprise - Databricks & Glean | BG2 Guest Interview (December 23, 2025) · at 22:44 ▸ · at 26:00 ▸
When this applies
when deciding where to position a business in a stack whose foundational layer is commoditizing
The takeaway
In a stack with one highly commoditized layer, expect value capture to concentrate at the layers that embed defensible lock-in (data, workflow, switching costs), even if disagreement exists on exactly which layer that is.
Why it works
Commoditized layers (the LLM/intelligence layer, per Ghodsi) can still support very valuable 'fab-like' companies, but defensible differentiation and switching costs concentrate in the application layer, which embeds proprietary data and workflow lock-in (akin to OS/platform stickiness vs. a commodity utility).
One level up
Track where switching costs and proprietary assets accumulate, not where raw technical capability originates.
Where it stops holding
This stops working if application-layer winners themselves get disrupted by new entrants (as internet portals were by Google/Facebook), so today's incumbents at any layer are not guaranteed permanence.
Detail
hierarchy:
- Application/app-platform layer (Ghodsi: 'most of the value will accrue to the apps')
- Intelligence layer (Arvind: 'maybe it'll capture half of the enterprise value', will be 'pretty thick')
- Data layer (foundational but necessary, not sufficient alone)
- Raw LLM commodity layer (lowest differentiation, though can still be 'TSMC-like' valuable fabs)
rationale: LLMs are interchangeable like a commodity ('you can get them, but that doesn't mean those companies are not gonna be valuable... they're gonna be kind of like these fab like companies'), while proprietary data and app experience resist commoditization.
In their words
I know we do think that the intelligence layer is actually gonna be a pretty thick one. Maybe it'll capture half of the enterprise value.
Commodity LLM, proprietary data as moat
Source
BG2Pod with Brad Gerstner and Bill Gurley — AI Enterprise - Databricks & Glean | BG2 Guest Interview (December 23, 2025) · at 6:45 ▸ · at 22:44 ▸
When this applies
when a core technology input becomes interchangeable across all competitors
The takeaway
When your core technology input commoditizes, value shifts to the proprietary assets layered on top of it.
Why it works
LLMs are now functionally interchangeable (like gas from any station), so competitive advantage cannot come from model choice; it must come from assets competitors cannot replicate, namely a company's proprietary data and unique business processes.
One level up
Build differentiation where imitation is structurally impossible, not where it is merely currently absent.
Where it stops holding
This stops working when a company has no genuinely unique data or process to leverage, in which case building on a commodity layer yields no edge and risks becoming 'demo-ware.'
Detail
procedure:
- Recognize that LLM quality differences are now too fast-moving to matter competitively
- Identify what data or process is unique to your company and inaccessible to competitors
- Build AI systems specifically to leverage that unique data/process, not generic commodity capability
- Avoid building 'commodity stuff' any company could replicate
rationale: There is no AI out there that understands your business processes and secret sauce; that's the actual moat.
In their words
It really comes down to that. Or if you have a complicated process that just your company has, this is how you deliver your product and services in your company. And it's, that portion can be disrupted with AI somehow. If you can do that, now you can get ahead of your competition.
Three camps of AI capital deployment
Source
BG2Pod with Brad Gerstner and Bill Gurley — AI Enterprise - Databricks & Glean | BG2 Guest Interview (December 23, 2025) · at 14:18 ▸ · at 17:56 ▸ · at 35:22 ▸
When this applies
when categorizing participants in a hyped emerging technology to judge which bets are speculative versus grounded
The takeaway
Segment participants in a hyped technology by their actual objective and capital exposure before judging whether 'the whole thing' is overvalued.
Why it works
Different participants in an AI ecosystem pursue different goals (superintelligence moonshot, fundamental research, applied enterprise value) with vastly different capital intensity and risk profiles; lumping them together misdiagnoses where bubble risk actually concentrates.
One level up
Avoid binary bubble/no-bubble judgments on an entire industry; risk concentration varies by sub-segment.
Where it stops holding
This stops working when the camps are so intertwined (shared capital, shared customers) that isolating risk to one camp doesn't protect the others from spillover.
Detail
procedure:
- Camp 1: Superintelligence quest (frontier labs, scaling-law bet on ever bigger clusters) — most capital-intensive, highest bubble risk
- Camp 2: Fundamental researchers (e.g., Rich Sutton, Yann LeCun) who doubt camp 1's approach and estimate ~20 years to a new paradigm — low capital, 'sober', likely correct but ignored
- Camp 3: Applied enterprise builders (Databricks, Glean) using existing AGI-level capability to extract real economic value — not capital-intensive, not in the speculative bubble
benchmarks: {"capex_context":"~$250B spent on NVIDIA chips implies ~$500B total capex requiring ~$1T AI revenue to justify, versus ~$400B total software industry revenue today"}
rationale: Bubble exists specifically in 'startups with zero revenue worth $10,20,30 billion,' not uniformly across the whole AI space.
In their words
And then there's the third camp, which is us trying to make this valuable. We're not in a bubble in a sense that we're not spending huge amounts of capital on what we are doing.
High failure rate signals healthy experimentation
Source
BG2Pod with Brad Gerstner and Bill Gurley — AI Enterprise - Databricks & Glean | BG2 Guest Interview (December 23, 2025) · at 0:00 ▸ · at 2:13 ▸
When this applies
when adopting any new, rapidly-evolving technology across an organization
The takeaway
A high experimental failure rate in an emerging technology is evidence of adequate risk-taking, not of the technology failing.
Why it works
If most experiments with a new technology succeed, you aren't pushing into genuinely new territory; a high failure rate indicates teams are stretching to find real use cases rather than playing it safe.
One level up
Judge innovation programs by whether they are exploring the frontier, not by their raw success rate.
Where it stops holding
This stops working when failures stem from poor execution or lack of basic engineering discipline rather than genuine exploration, or when failure rates stay high for years without any conversion to success.
Detail
rule: Treat a ~95% failure rate in new-technology pilots as a sign of appropriate experimentation volume, not alarm.
threshold: 95% failure cited as acceptable/expected in year one of a new technology cycle
rationale: You want everyone in the industry eager to experiment; low failure rates would mean not trying enough.
In their words
You hear these 95% of projects fail, but, like, you know, like, that's actually what you want.
Keyboards will disappear as speech interfaces finally mature
Source
BG2Pod with Brad Gerstner and Bill Gurley — AI Enterprise - Databricks & Glean | BG2 Guest Interview (December 23, 2025) · at 39:56 ▸
When this applies
when judging whether a long-promised interface technology (e.g., voice) is actually near a breakthrough versus still immature
The takeaway
Use continued reliance on the old interface as the real test of whether a new interface technology has actually matured, rather than trusting surface-level hype.
Why it works
Ghodsi argues the persistence of keyboard use is itself evidence that speech interaction hasn't truly been 'nailed' yet, despite perceptions of progress; he predicts the remaining gap will close soon.
Where it stops holding
This prediction fails if speech interfaces remain inadequate for precision, privacy, or noisy-environment use cases where keyboards retain structural advantages.
Detail
rule: If people still default to keyboards, speech interaction is not yet solved, regardless of perceived progress.
threshold: 'As long as you're using your keyboard, we haven't nailed speech'
rationale: I think we're this close to completely eliminating keyboards.
In their words
We haven't actually nailed speech. I know it feels like we have, but we haven't because you're still using your keyboard. So as long as you're using your keyboard, we haven't nailed speech. But I think we're this close to completely eliminating keyboards.