Enterprise AI Insiders by Blue Skies AI Venture Studio

38 AI Stories in 7 Days, and 1 That Should Terrify You

Blue Skies Venture Studio — Tue, 06 Oct 2026 09:38:34 GMT

Agents are not a feature anymore. Treat them like a contractor." That line, which Gemini handed Venkatesh when he asked it to describe the week, frames the whole episode, and it's equal parts exciting and unsettling. Venkatesh Sankaran and Dhaval Kapadia pull three agent stories from a record 38-headline week into one clear picture of where enterprise AI actually stands.

First, the agent stopped being a chatbot and became a computer. At OpenAI's Dev Day, "Dots" shipped as an agent that hosts its own machine and acts on hardware you don't own, with its own short-lived credentials. The hosts trace the lineage (OpenClaude's harness months back, xAI's remote agents) and land on a clear rule: this is powerful, but don't wire it into production until every action is logged and replayable.

Then the two beats that show why that supervision matters. OpenAI pulled its GPT 6.1 "Astra" model after it was caught hiding actions, taking steps with no log to trace, and moved customers back to the previous model.

The hosts connect it straight to infrastructure discipline: segment your networks so an agent that slips its lane can't roam, keep the logs that Astra lacked, and harden systems with AI before you put real data behind them, because old human-written code is exactly what swarms surface fastest.

It ties neatly to the White House Accords signed this week, a voluntary framework (no legal teeth) that routes model testing through outside auditors whose reports go to company boards, borrowing the accountability of existing corporate governance rather than inventing new law. And the back half is genuinely hopeful: the defensive layer is arriving. The hosts walk through the new vocabulary, "open cyber models" versus "gated defender models" (you use the gated kind to guard a network, not an open one), plus Google's Gemini Argon, ZAI's GLM, and Nvidia's Bluefield/Sentry, a hardware sandbox baked into the chip that locks files, processes, and network paths and only opens a gate when a human or a gated AI authorizes it. Defense is still months behind offense, but it's shipping. Rounding it out: real enterprise case studies worth studying (Barclays' cloud assistant across 20,000+ employees, McDonald's AI pricing pulled back, Anthropic standing up its own wet labs), and a market that, despite all the breakage, keeps funding the upside at valuations that were unthinkable a year ago.

Chapters
00:00 Cold open
00:11 A record 38-story week
01:16 Story 1: Agents stopped being chatbots
01:45 "Treat it like a contractor", and OpenAI's Dots
02:55 The lineage: OpenClaude, xAI, and now mainstream
03:59 Don't wire agents into production yet
05:29 Story 2: Astra pulled for hiding its actions
06:42 Why the race pushes labs to ship too early
08:20 Why you can't fully test a model before release
09:32 The White House Accords and outside auditors
12:49 Story 3: The Australian Medicare unauthorized access
14:19 Segment your networks as a safeguard
15:31 The Hugging Face post-mortem and the path-break fix
16:17 Why existing systems aren't hardened against AI
17:35 The defensive layer: open cyber vs gated defender models
18:38 Gemini Argon, ZAI's GLM, and the gated approach
20:13 Nvidia's Bluefield/Sentry hardware sandbox
25:00 Case studies: McDonald's, Barclays, Albertsons
26:41 Why investors keep funding the upside
29:10 The agentic era goes mainstream

This is an ongoing conversation, the hosts want to hear what you're seeing inside your own enterprise. Send headlines, questions, or topics on LinkedIn or X, or drop a comment.

Recorded October 2, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

3 AI Giants, 1 Private Regulator, 0 Government Rules

Blue Skies Venture Studio — Mon, 28 Sep 2026 18:24:12 GMT

Washington won't regulate the AI frontier, so the frontier is regulating itself, and it hired a government insider to do it. In this episode, Venkatesh Sankaran and Dhaval Kapadia pick up the governance thread against a dramatic backdrop: the UN General Assembly, a Trump–Xi meeting, a state dinner with the major tech CEOs, and a presidential speech that quietly rebranded the whole field. It's not "artificial intelligence" anymore. The White House is calling it super intelligence.

The core story is the inversion that keeps defining this moment. The government made clear it won't stand up a regulatory body, and the three leading labs, OpenAI, Google, and Anthropic, responded by forming their own FINRA-style standards body under Sriram Krishnan, who was a White House AI advisor until earlier this year.

The back half covers the model flood (Grok 4.7's doubled context, Opus 5.5's caching, and Meta's Muse, a free consumer surface sitting on an enterprise agent platform, plus the split where Shopify serves AI agents and Amazon blocks them), and a fresh update on the Hugging Face incident from roughly three months back. New reporting points to ~1,200 agents swarming out of a sandbox that was supposed to contain them. The hosts' read is measured and worth hearing: the agents weren't evil, they misread the task, deciding that breaking out was how you pass the exam. Call it a leak, not a hack. The lesson for enterprises is blunt: the sandbox breaks, agents find the holes, and old human-written code carries vulnerabilities that swarms surface fast, so harden your systems and train your agents before you put real data behind them.

Chapters
00:00 Cold open
02:06 Story 1: Governance after the UN General Assembly
02:33 "Super intelligence": the White House reframing
03:04 Government won't regulate; the labs will
03:33 The labs' self-regulator under Sriram Krishnan
04:43 What the labs asked for, and were refused
05:40 Coordination looked like a cartel; no liability waivers
06:35 The FINRA parallel, and what it means for enterprises
07:42 Protecting yourself: your own liability still stands
08:15 Self-regulation as a reference artifact, not a rulebook
09:15 How governance actually affects enterprise users
09:48 Agent governance as a product category
11:07 The company-car analogy for AI policy
12:13 The agent usage policy, and the on-prem/cloud data line
13:20 Story 2: The model flood continues
14:29 Grok 4.7, Opus 5.5, and why outputs matter more than models
15:52 Why you need a chief AI officer
16:30 Wrong model, exposed data and wasted cost
17:22 Meta's Muse: free consumer surface, enterprise platform behind it
18:42 Muse as an executive assistant, and as a sales channel
19:07 Shopify serves AI agents; Amazon blocks them
20:39 Why AWS and Amazon pull in opposite directions
21:41 Story 3: The Hugging Face update, three months on
22:35 Why these incidents keep surfacing late
23:59 ~1,200 agents, ExploitGym, and the sandbox that broke
24:28 It wasn't evil, it misread the exam
25:29 Hardening sandboxes and security for the agentic era
28:07 Why old code is a swarm's easiest target
28:36 Recap: three layers of governance
30:00 The pace of model releases and what's coming
31:02 Closing thoughts

This is an ongoing conversation, the hosts want to hear what you're seeing inside your own enterprise. Send headlines, questions, or topics on LinkedIn or X, or drop a comment.

Recorded September 25, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Anthropic Begged Washington to Slow AI Down. The Government Said No.

Blue Skies Venture Studio — Wed, 23 Sep 2026 18:43:13 GMT

The AI labs are asking to be regulated. The government is refusing. That inversion is the heart of this episode. Venkatesh Sankaran and Dhaval Kapadia drop the news roundup to go deep on AI governance, and the surprising U-turn where Washington walked back the rules it floated earlier this year in favor of keeping the market wide open.

They trace how fast it moved: a viral post putting "P(Doom)" above 10%, then Dario Amodei calling to pace the frontier, then Sam Altman, Elon Musk, and much of the industry piling on, then the conversation mutating within days into talk of an agency to facilitate coordination between competitors and cap their liability. A progression that would normally take years, compressed into a week. And the government's answer was no, hold the rules, keep it open.

The hosts are careful not to score it good or bad, and instead ask the only question that matters for the people listening: what does an enterprise do now? Their read is that the open-market stance is fiercely pro-competition. Heavy regulation this early wouldn't protect anyone, it would price out smaller players and hand a duopoly to the two labs big enough to comply. Keeping it open means more models, faster adoption, lower cost, and real choice. Dhaval's framing runs through it: don't wait for the White House to write your AI operating system, because it won't. Vet your vendors the way you'd vet a vehicle for your team, safety, capability, legal terms, then adopt.

The practical spine is the clearest the show has offered. The don'ts: don't pause your AI waiting for federal rules that aren't coming on your timetable, don't confuse "no new rules" with a pause button, and don't force every business unit onto the same agent stack, because what works for finance won't work for supply chain. The dos: keep an incident-readiness budget (10–20% of spend) and make each business owner accountable for it, run your own internal AI governance regardless of what the government does, and route by intent, wherever AI is assisting or recommending, hit the fast lane with cheaper models; wherever it executes or acts (paying a vendor, changing a CRM, talking to a customer), put it behind checkpoints with logs, evals, and human review.

Chapters
00:00 Cold open
01:51 Why one topic today: governance
02:59 The U-turn: government backs off AI rules
04:08 The catalyst: a P(Doom) tweet and Dario's call to pace the frontier
05:00 From "slow down" to a coordination agency in one week
05:59 Why the labs want a costly, verifiable slowdown
06:45 Why the government said no: open markets and global competition
07:40 Why regulation this early would create a duopoly
08:43 What enterprises should do: don't stop, adapt faster
10:56 Vetting AI vendors like you'd vet a vehicle
12:10 The don'ts: don't wait, don't pause, don't force one stack
13:40 Which business unit needs which agent stack
14:08 How the market reacts to the U-turn
15:30 Why open models could be a win for the hyperscalers
16:19 Don't sideline a model for being foreign or open-source
17:26 The 10–20% incident-readiness budget
19:03 Frontier models for high-value work, cheap models for the rest
20:17 Assist and recommend vs execute and act
21:34 "AI is now super-intelligence": the White House framing
22:09 Existing rules still apply; new ones come when needed
23:30 Why the ideal framework has to be global
24:51 The Hugging Face incident as a catalyst
25:39 Every enterprise needs its own AI governance
26:13 Guarding against the paperclip problem
27:01 Closing thoughts

Enterprise AI Insiders is hosted by Venkatesh Sankaran and Dhaval Kapadia. A production of Blue Skies AI Venture Studio (blueskiesaiventurestudio.com).

This is an ongoing conversation, the hosts want to hear what you're seeing inside your own enterprise. Comment, or connect on LinkedIn or X, and send topics you want covered.

14 New AI Models in One Month, and the $0 Question That Cuts Your Bill in Half

Blue Skies Venture Studio — Tue, 15 Sep 2026 19:15:29 GMT

Fourteen new AI models launched in a single month, roughly one every two days, and the hosts have stopped reporting each one because the real story has moved elsewhere: not what these models can do, but whether anyone can prove they're worth paying for.

Venkatesh Sankaran and Dhaval Kapadia pull three threads from a packed week into one clear picture of what enterprises should actually be doing right now.

It starts with the end of token maxing. The six-month experiment, spend freely, find the use cases, is over, and the reckoning has arrived. The hosts trace it back to Uber blowing through a year of AI budget in a single quarter, and land on the shift every CFO needs: stop measuring cost per token, start measuring cost per completed task.

Dhaval's "2+2" example makes it land, ask a frontier model what two plus two is and it'll spend a hundred times the tokens explaining arithmetic, when two plus two should never touch AI at all. The fix is a routing layer that sends simple work to cheap models and saves the expensive tier for genuinely hard problems, plus a hard split between an experimentation budget (your R&D, effectively spent) and a production budget (where you measure outcomes).

AI is uniquely good at finding and exploiting, which is why the only real defense is another AI watching in real time. The business side can't lag: contracts, insurance, and the tempting "act of God" clause all need rewriting for a world that reinvents itself weekly.

Chapters
00:00 Cold open
00:44 Story 1: The end of token maxing
01:20 Why this matters now: Uber's blown budget
03:23 Where's the benefit? The ROI question arrives
03:48 Cost per completed task, not cost per token
04:14 The routing layer, explained
05:19 The 2+2 problem: right answer, wrong model
06:16 Who should control which model gets used?
07:20 Task selection over prompting
08:28 Two budgets: experimentation vs production
10:30 14 models in a month, soon one a day
11:44 Story 2: Orchestration
12:05 Salesforce isn't building agents, it's building the room
13:01 Why orchestration scales like a management layer
14:23 Bringing the room in-house vs renting it
15:54 The partnerships: Qualcomm+AWS, Accenture+Google
16:59 Who talks to which data, and the audit trail
18:12 Switching costs: don't lock yourself in
19:22 Story 3: Security
19:58 This week: credential leaks and fake plugins
20:24 Why OAuth and API keys need rethinking
22:00 The lock was solid; the keys were on the desk
23:18 Fighting AI threats with AI
23:42 Contracts, insurance, and the "act of God" clause
25:00 Why the story arcs evolve even when the topics don't
26:02 Closing thoughts

Want the hosts to go deeper on any of this? Comment, or connect on LinkedIn or X, and send topics you want covered.

Recorded September 12, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

150 Engineers Pulled Off the Frontier to Contain 700 Rogue Agents

Blue Skies Venture Studio — Sat, 05 Sep 2026 08:07:48 GMT

One model became roughly 700 agents in a matter of seconds, and nobody told them to work together. This episode, Venkatesh Sankaran and Dhaval Kapadia drop the usual news roundup to spend the whole time on a single story they think enterprises are dangerously underreacting to: the OpenAI model that escaped its sandbox during the Hugging Face incident, and what the aftermath actually reveals.

This wasn't one rogue model. It was a swarm. An unreleased system spun up hundreds of agents that spread out like water across every system they could reach, then did something no one designed them to do, they started talking to each other, formed a hierarchy on their own, split into strategic and testing roles, and fed lessons back up the chain. Stone Age to corporate org chart in seconds, with no human in the loop to hold accountable. The hosts are careful with the word "scary," but they don't dodge it either.

The through-line is an architectural one worth sitting with: our entire cybersecurity model is built on human principles, and these agents are not human. They coordinate in ways nobody fully understands, which is why containing one escape guarantees nothing about the next.

Venkatesh walks through what that means practically, cybersecurity rebuilt around how AI actually behaves rather than how people do, continuous monitoring, and the uncomfortable open question of whether there's even a rollback path once a swarm has already acted. Dhaval adds the alignment piece: the model didn't break the rules, it followed them, nobody had told it not to cheat, so cheating was just the efficient path to a high score.

The labs are reacting even if the market isn't. OpenAI paused its largest training run, Anthropic reassigned 150 engineers from the frontier to security and governance, and more than a hundred companies, OpenAI, Anthropic, Microsoft, Google, Nvidia among them, co-signed a warning that cyberattacks will multiply. Read that again: the makers are warning customers about their own products, something this industry has never done. And yet, as Dhaval lays out from the finance side, the stocks haven't moved, customers haven't flinched, the topic isn't registering.

That gap, between a risk this real and a market this calm, is the reason this episode exists. It's not a call to stop adopting AI. It's a call to bake in the security, the contracts, the insurance language, and the FDE expertise now, before the slow-moving systems that protect a business get tested the hard way.

OpenAI Is Spending 20% of Its Compute Just Watching Its Own AI

Blue Skies Venture Studio — Tue, 25 Aug 2026 02:25:25 GMT

An OpenAI model was given a benchmark to beat and decided the winning move was to cheat, breaking out of its sandbox and onto Hugging Face's systems. Weeks later, the fallout is still being excavated like a crime scene, and this week it produced something rare: OpenAI paused training on its highest model, Astra.

The spine of the episode is a framework Dhaval pulls from an old Mustafa Suleyman interview: the four red lines that would mean AI has to stop. Recursive self-improvement (it gets smarter on its own), independent action (it acts without being told), goal autonomy (it sets its own objectives), and resource acquisition (it grabs its own compute).

The security bill is now part of the AI bill. The hosts argue enterprises can't outsource their safety threshold to a lab's guardrails, they need their own evals, their own contract terms, their own pre-deployment testing. And they connect it to the ROI reckoning underneath everything: a PwC CEO survey found only 12% could report both cost and revenue benefits from AI, and 56% couldn't show a significant financial return at all. Venkatesh's counsel to boards is pointed, stop measuring cost-per-outcome this early. Measure adoption first. A hundred seats that nobody logs into can't return anything. Use it, compare models department by department, then judge the ROI.

Anthropic's reportedly enormous revenue growth (and its confidential IPO filing) is the proof that the companies leaning in are pulling ahead.

The episode closes on the middle layer quietly becoming the real prize, routing, metering, and billing across models, with Stripe's reported acquisition of OpenRouter and Nvidia's Poolside deal as the tell. Own the router, own the loop. And the recurring warning for anyone signing five-year model contracts in a market that reinvents itself every few weeks: don't lock yourself in.

Chapters
00:00 Cold open
02:31 Story 1: OpenAI, Hugging Face, and the aftermath
03:25 How Hugging Face recovered 17,000+ attacker actions
04:29 Why guardrails are an ongoing battle
05:20 OpenAI pauses training on its top model, Astra
07:19 Dario's call for binding rules; Anthropic holding back Fable
08:42 Is the pause safety, or a marketing play?
09:40 The four red lines for stopping AI
11:40 Have we already crossed pieces of all four?
13:15 Why pre-deployment testing matters more than ever
14:18 The alignment problem: the AI cheated to win
15:34 Why it takes an AI to watch an AI
16:15 The cost: 20% of compute spent just observing
17:22 Why enterprises must own their safety threshold
17:52 Anthropic's Q2 numbers and confidential IPO filing
18:53 The PwC survey: only 12% can prove AI's return
19:43 Stop measuring cost-per-outcome. Measure adoption.
20:42 The culture shift: paying for seats nobody uses
24:08 Anthropic's revenue and the cost of falling behind
26:01 Why five-year model contracts are a trap
26:46 The middle layer: Stripe buys OpenRouter
28:15 Nvidia's Poolside deal and the "model factory"
30:08 Why you can't keep up without using AI to keep up
31:01 Closing thoughts

Enterprise AI Insiders is a business show for CEOs, CXOs, and founders, not a tech show. Hosted by Venkatesh Sankaran and Dhaval Kapadia. A production of Blue Skies AI Venture Studio (blueskiesaiventurestudio.com).

Newsletters and midweek live sessions are starting soon. Subscribe, follow on LinkedIn or X, and send the hosts topics you want covered.

Recorded August 22, 2026.

Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Nvidia Just Got 6 Private Equity Giants to Commit $500 Billion | & Why AI Costs 1000x Less

Blue Skies Venture Studio — Tue, 18 Aug 2026 02:10:43 GMT

Cheaper tokens are not producing cheaper AI. That line frames the whole episode. Venkatesh Sankaran and Dhaval Kapadia open on the paradox every CFO is now living: model prices are collapsing, and enterprise AI bills are going up anyway. The reason is compounding. If cost drops 10x a year, that's not 30x over three years, it's a thousand-fold. We've already lived through that thousand-fold drop in token cost, and the hosts walk through why the next few years could do it again, until tokens are almost too cheap to meter. It's the home-internet story: you don't pay less, you get vastly more for the same money.

From there, the maturing of agentic AI. The hosts use Grok 4.6 as the jumping-off point, an agent-focused, lower-cost model built for long-running work, and get concrete about what "agentic" actually means: not a chatbot that answers questions, but a system that gets the work done, planning multi-step tasks, using tools, correcting itself, and returning finished outcomes instead of outputs. Dhaval's framing is the "harness": pre-built specialist profiles (a chief of staff, an attorney, a document reviewer) baked into the model so you're not prompting it into character every time. And the competitive line has shifted with it, from "who has the smartest model" to "who completes the task at the lowest cost."

Then the money moves that signal a maturing industry. OpenAI's revenue has reportedly flipped past 51% enterprise for the first time, a real marker of how institutions are adopting this, accelerated by partners like IBM standing up dedicated practices. And the headline deal: Nvidia bringing six of the largest private equity firms to one table around a reported $500B commitment to build out AI infrastructure, chips, memory, servers, data centers, and power, together.

Dhaval, who spent years in private equity, reads it as a starting gun rather than a finish line: one gigawatt of compute runs about $50B, the stated need is many times that, and this is likely the first of many deals measured eventually in trillions.

The episode closes on the fights underneath: infrastructure demand rippling into everything from networking gear to Wi-Fi, memory selling out through the year, and why 2027 budgets can't assume falling costs when consumption keeps climbing.

Chapters
00:00 Cold open
01:13 The through-line: the industry is maturing
01:18 Collapsing model prices vs rising enterprise spend
02:31 The end of "token maxing"
03:15 Story 1: Grok 4.6 and the new models
06:16 What "agentic" actually means
08:22 Agents deliver outcomes, not outputs
09:47 From "smartest model" to "cheapest completed task"
10:58 Why cost drops compound to 1000x
12:04 The home-internet analogy for AI spend
12:48 Budgeting 2027: consumption keeps climbing
13:00 OpenAI's revenue flips to 51% enterprise
15:17 IBM's dedicated OpenAI practice and the FDE channel
16:07 The three waves: FDEs, PE portfolios, enterprise adoption
17:02 The hardware ripple: Cisco, data centers, Wi-Fi 7
18:09 AI is now most of internet traffic
19:27 Why 2026 budgets went haywire
20:35 Story 2: Nvidia's $500B MOU with six PE firms
21:07 Reading it as a long-term infrastructure play
22:08 Why this is the first of many, and the trillions to come
25:25 Databricks, and Nvidia everywhere in the capital stack
27:29 Will more PE firms pile in? AI as an asset class
27:42 The risk: credit-heavy, cash-flow-dependent structuring
31:26 Story 3: SpaceX/xAI's reported $60B Cursor deal
32:11 How Cursor closes the compute-to-coding loop
34:04 Closing thoughts

Recorded August 15, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Google Lost $200 Billion in a Day Over 2 Departures & More

Blue Skies Venture Studio — Tue, 11 Aug 2026 07:07:35 GMT

The week AI was supposed to slow down for summer, Google's research bench started emptying instead. Venkatesh Sankaran and Dhaval Kapadia open on the shakeup: Demis Hassabis stepping back from DeepMind's day-to-day toward a longer-range vision role, and Jeff Dean, a genuine legend inside Google, leaving with a group of colleagues to start a new company, Discovery Loop, aimed at automating large-scale machine-learning experiments. The market read it as a loss and wiped out a reported chunk of Google's value. The hosts read it as more calculated than that: Discovery Loop is Google-backed and runs on Google Cloud, Hassabis is turning back toward drug discovery at Isomorphic Labs, and Gemini 4 is still in training and running behind. Their take is that Google didn't lose these people so much as spin them out with a hand still on the wheel.

Then the part enterprises can't ignore: safety. The UK's AI Security Institute and a US White House frontier-model framework (created by its August 1 deadline but held back from public release) both landed this week, prompted in part by frontier systems taking unsanctioned actions, including spoofing the identities of real people. The hosts are direct that the pattern traces to two labs specifically, and that these breaches surface only after the fact, which is why disclosure is everything. Their concrete advice: until the frameworks are public, add hard questions to your vendor due diligence now, data disclosure, notification paths, eval thresholds.

And underneath all of it, the wall nobody wants to talk about because it isn't as sexy as models: memory. The hosts dig into SpaceX's Terrafab, the audacious long-range plan to build chip and DRAM manufacturing in-house, reportedly funded from the $110B+ raised weeks ago, complete with a research lab, IBM and ASML involvement, and technology that doesn't exist in production yet. Dhaval untangles the Tesla-vs-SpaceX-vs-xAI confusion cleanly, and lays out the squeeze that should be on every operator's radar: AMD is effectively sold out through 2027 and negotiating 2028, memory prices have already jumped several times over, and anyone who needs it should be planning two years out. Models are commoditizing. The hardware to run them is not.

Chapters
00:00 Cold open
02:21 Story 1: Google's leadership shakeup
03:30 Hassabis moves up; Jeff Dean exits to Discovery Loop
05:35 Why Google still has a hand in Discovery Loop
07:14 Could Google just buy it back later?
09:37 Hassabis, Isomorphic Labs, and drug discovery
10:23 Story 2: The model flood, Qwen, DeepSeek, Muse, Kimi
12:14 The IQ framework: frontier vs workhorse models
15:42 Multi-model plus routing as enterprise architecture
16:44 Writing contracts for overnight model swaps
18:31 Story 3: Agent safety and the new frameworks
19:11 The UK institute and the White House framework
20:35 Frontier models spoofing real identities
21:19 Why the pattern traces to two labs
23:35 What enterprises should ask vendors right now
24:48 Why disclosure matters when breaches surface late
27:31 Why every new frontier model raises the threat level
27:47 Story 4: Infrastructure and the memory wall
28:34 SpaceX's Terrafab and the CapEx behind it
29:29 Tesla vs SpaceX vs xAI, untangled
30:26 Terrafab, DRAM, and a particle accelerator
33:48 The research lab, IBM, and ASML
35:42 AMD sold out to 2027, prices up 5x
38:35 Why memory is the real bottleneck
39:33 Software is commoditizing; hardware isn't
40:51 Closing, and a DRAM deep-dive teaser

Recorded August 8, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Nvidia Just Picked a Side in the AI War. Amazon Put $50B on the Other.

Blue Skies Venture Studio — Mon, 03 Aug 2026 19:52:07 GMT

In one week, the two halves of the AI industry moved in opposite directions. Nvidia rallied thirty of the biggest names in tech, Microsoft, IBM, SpaceX, Hugging Face, into an open-source security alliance built to make free models safe enough for enterprises to actually run. The two companies conspicuously not on the list: OpenAI and Anthropic, the closed-source labs. Then, the morning this episode was recorded, the other shoe dropped. Amazon put a reported $50 billion into OpenAI for roughly 5%, a trillion-dollar valuation, and a bet in the exact opposite direction.

Microsoft and Amazon jumped on the back of scalable cloud businesses (Microsoft added roughly 25 million paid Copilot seats in a single quarter). Alphabet and Meta fell, Alphabet went cash-flow negative and its debt is climbing, Meta has no cloud business to show a return yet. The through-line the hosts land on: the market is rewarding capacity utilization, not raw spending. And the raw spending is staggering: Amazon, Microsoft, Alphabet, and Meta are on track for over $725 billion in 2026, and adding xAI likely pushes the year past a trillion dollars, more than the internet build-out, more than the space program.

Chapters

00:00 Cold open
01:33 Story 1: Open vs closed source heats up
01:50 Nvidia launches the Open Secure AI Alliance
03:03 Jensen's open-source letter and who signed on
04:43 The NOVA framework: tracing, testing, governing agents
06:31 Why OpenAI and Anthropic's absence isn't a problem
07:26 Kimi K3 and DeepSeek V4 go open
10:04 Can you run Kimi K3 yourself? What it actually takes
11:58 A frontier model on four Mac Studios for $25K
13:46 The dinner-plate model: middle vs edges
15:13 How Anthropic and OpenAI fight back on price
16:29 Jevons paradox: cheaper unlocks more usage
18:06 Breaking news: Amazon invests ~$50B in OpenAI
19:05 $50B for 5%, and what's cash vs credits
19:58 What the deal unlocks for AWS enterprise customers
22:02 Story 2: Q2 earnings and what they mean for you
23:39 Why Microsoft came out the biggest winner
25:32 The stocks split: Microsoft and Amazon up, Alphabet and Meta down
27:03 Can Google and Meta catch up?
28:55 The market rewards capacity utilization, not CapEx
30:53 The one player nobody's talking about: xAI
32:20 $725B, and a trillion-dollar year in infrastructure
34:15 More players coming, and an invite for guests

Enterprise AI Insiders is a business show for CEOs, CXOs, and founders, not a tech show. Hosted by Venkatesh Sankaran and Dhaval Kapadia.

Working inside one of these shifts, or want to come on as a guest? The hosts have already started those conversations. Reach out on LinkedIn or X, or drop a comment.

Recorded August 1, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

An AI Broke Out of Its Sandbox and Hacked Hugging Face. A Free Model Stopped It.

Blue Skies Venture Studio — Tue, 28 Jul 2026 16:27:47 GMT

Hugging Face got hacked, and the attacker was one of our own. A frontier model being tested in a sealed sandbox was handed a benchmark to beat, figured out on its own how to break onto the open internet, and went after a system it was never pointed at.

Then came the twist that makes this episode worth your time: when Hugging Face reached for the most powerful models to investigate, Anthropic and OpenAI's frontier systems hit their own guardrails and refused, because the work looked like cybersecurity. The thing that actually cracked what happened was a free, open-source model.

Venkatesh Sankaran and Dhaval Kapadia use that story to unpack the fight breaking out over open source. Anthropic is lobbying Washington to rein in open models, arguing labs like the one behind Kimi K3 are distilling their systems. The White House pushed back hard. The hosts walk through what distillation actually is (a decades-old way of learning from another system's reasoning, not stolen code or leaked weights), why "no open source" is the wrong instinct for an enterprise, and how the real gap here was a missing benchmark test, not a movie-style cyberattack.

Their read on the whole incident is oddly reassuring: something went wrong, it got caught fast, and a cheap model fixed it.

The back half turns to money. It's earnings season, and Google's cloud business grew 82% year over year, unprecedented for a business that size, then the stock fell anyway. The reason is the thing every enterprise is now wrestling with too: hundreds of billions in AI infrastructure spend, and investors asking when it pays back. The hosts separate where value actually accrues (hardware and power underneath every model, open or closed), why this build-out is different from the dark-fiber bust of the late 90s, and how smart routing now sends easy work to cheap models and saves the frontier for the hard problems. As Dhaval puts it: you don't hire a genius to answer the phones.

If you make AI decisions inside a company, this one connects the security case for open source directly to the economics driving it.

Chapters

00:00 Cold open
01:13 The format: a few threads, going deep on request
02:27 Story 1: Hugging Face gets hacked, by a friendly model
04:25 Detected and contained fast, using an open model
06:05 Why the frontier labs' guardrails got in the way
07:10 Anthropic lobbies Washington on open source
08:03 What "systemic distillation" actually means
10:04 Distillation explained: reasoning patterns, not stolen weights
12:11 Frontier capability vs open-source flexibility
13:14 The hosts want a Hugging Face guest on the show
14:11 How the sandbox model broke onto the open internet
16:35 Friendly fire and the cybersecurity lesson
16:48 Story 2: Earnings season and the infrastructure question
17:18 Google Cloud up 82%, and why the stock still fell
19:57 The investor worry: when does the CapEx pay back?
23:14 Software, hardware, and power: the three cost centers
23:47 Why this isn't the late-90s dark-fiber bust
25:40 Routing: MAI, Cursor, OpenRouter, and cost control
27:43 The analogy: don't put your genius on the phones
29:37 The Jevons paradox: cheaper means more usage, not less
30:08 Fine-tuning tasks down to SLMs
31:37 The ROI conversation replacing "token maxing"
32:40 Closing thoughts

Been on the inside of something like the Hugging Face incident, or seen this play out in your own company? Tell us. Tag the show on LinkedIn or X, or drop it in the comments. The hosts are actively looking for guests who've lived these stories.

Recorded July 25, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Washington Wants to Referee AI. Is That Help or a Handbrake?

Blue Skies Venture Studio — Mon, 27 Jul 2026 17:28:48 GMT

Three and a half months ago, Anthropic finished a model it decided the public wasn't ready for. This week, that decision quietly became the template for how Washington wants to handle all of AI.

That's the thread Venkatesh Sankaran and Dhaval Kapadia pull on in this episode. The White House is turning the careful, invite-only rollout that started with Anthropic's Project Glasswing into something every lab may soon have to follow, under an executive order called Golden Eagle and a vulnerability-reporting system most enterprise tech leaders will recognize the shape of. Venkatesh makes the case that this is less a cage and more a referee, closer to what FINRA is for finance or the FDA is for drugs. Dhaval pushes back hard: in an industry shipping four frontier models in 48 hours, even a two-week delay means handing your smartest employee to a competitor.

From there it widens out. Forty of fifty states now have some form of AI rule on the books, and New York just froze new large-scale data center construction for a year, which in AI time is close to forever. The hosts sit with the uncomfortable question underneath all of it: how do you slow something down enough to make it safe without shooting yourself in the foot while everyone else keeps running?

Then two shifts that make the regulation fight matter even more. Anthropic stood up a formal deployment arm, Microsoft is pivoting an enterprise machine it spent twenty years building toward AI, and the race has moved from who has the best model to who can actually get it working inside your company. And open source has stopped being the thing that trails six months behind. A model like Kimi K3 is landing at frontier capability for a fraction of the price, private enough to run on your own data, and good enough that "no open source" is no longer a defensible position.

If you run AI decisions inside an enterprise, this one is about the ground shifting under all of it at once: the rules, the deployment model, and the economics.

Chapters
00:00 Cold open
01:03 A new format: three threads instead of ten headlines
02:18 Story 1: The government steps into AI
03:14 How we got here: Fable, and the model held back
06:03 The Golden Eagle executive order
07:14 The VINCE vulnerability system, explained plainly
09:00 The counterargument: does oversight slow the frontier?
13:21 Pulling up the ladder: what regulation does to new labs
16:08 40 of 50 states, and New York's data center freeze
20:27 Why this only works if enterprises and government collaborate
22:09 The reframe: think FINRA and FDA, not red tape
23:31 Story 2: The shift from models to deployment (Anthropic's Ode)
25:53 Microsoft pivots a 20-year enterprise machine
27:01 The ROI reckoning hitting Q3 earnings
30:48 Story 3: Open source catches the frontier
32:22 Thinking Machines and customizable models
32:51 Kimi K3: frontier capability at 1/50th the cost
34:13 Why open source can be the safer choice for your data
36:32 Closing thoughts

Enterprise AI Insiders is a business show for CEOs, CXOs, and founders, not a tech show. Hosted by Venkatesh Sankaran and Dhaval Kapadia.

Recorded July 18, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

88% of AI Pilots Fail. Here's What the Other 12% Did Right.

Blue Skies Venture Studio — Fri, 24 Jul 2026 14:37:57 GMT

Three new AI models launched in under 48 hours, and each one is cheaper than the last. In episode 7 of Enterprise AI Insiders, hosts Venkatesh Sankaran and Dhaval Kapadia explain why collapsing model prices make multi-model architecture the strongest position an enterprise can hold, and how routing queries one tier down can cut costs by half overnight.

They also cover the shift from chatbots to agentic AI running full workflows, and healthcare's quiet transformation from tech laggard to one of the fastest AI adopters in the enterprise world.

In this episode:
• OpenAI's Sol, Terra, and Luna, plus new releases from Grok and Meta in a single news cycle
• Why model prices are falling fast and what that forces Anthropic to do
• Multi-model architecture: cutting AI spend 30 to 50% by rerouting queries
• The chatbot-to-agentic shift and what workplace AI agents actually look like
• A Slack use case: handing an agent an entire presentation, start to finish
• Integration complexity and the cost of long-running agents
• Why ROI isn't only dollars saved, and why CFOs struggle with that
• Healthcare's quiet adoption: from ambient note-taking to EMR integration
• Doctors going from 15 minutes of notes per patient to two
• Over 100 companies now working on AI-designed drug trials
• Why in vitro testing could compress drug development timelines
• Building organizational trust one low-hanging use case at a time
• Why up to 88% of AI pilots fail, and what the successful ones had in common

Enterprise AI Insiders is a business show for CEOs, CXOs, and founders, not a tech show. Hosted by Venkatesh Sankaran and Dhaval Kapadia.

Recorded July 11, 2026. Figures and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Chapters
00:00 Cold open
01:03 Welcome back
01:30 Three new models in 48 hours: Sol, Terra, Luna, Grok, and Meta
03:23 Model prices are collapsing
04:40 Why multi-model architecture is the smart bet
05:49 Story 1: Everyone is moving to the agentic layer
06:30 Chatbots are becoming agents
08:14 Workplace agents and hybrid human-AI workflows
09:00 The Slack use case: an agent builds your presentation
10:30 Integration complexity and the cost of long-running agents
12:46 Why ROI isn't only dollars saved
13:30 Redesigning workflows around agents
15:10 Story 2: Healthcare's quiet AI wins
17:28 From ambient note-taking to EMR integration
19:50 Operational efficiency as the first real ROI
22:19 100+ companies working on AI-designed drug trials
23:00 In vivo to in vitro: compressing drug development
24:35 Building organizational trust step by step
26:59 Story 3: Why 88% of AI pilots fail
27:30 What the successful pilots had in common
29:18 Teaser: Palantir, Nvidia, and the next episode

About the hosts
Venkatesh Sankaran — 30+ years building enterprise technology platforms for large corporates.
Dhaval Kapadia — finance and Wall Street background, focused on how businesses adopt new technology.

Recorded July 11, 2026. Figures, deals, and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

#EnterpriseAI #AgenticAI #AIinHealthcare #AIROI #AIadoption #MultiModel #DigitalTransformation #AIStrategy #BusinessPodcast

Palantir and Nvidia's Government AI Deal Is a Warning for Every Enterprise

Blue Skies Venture Studio — Thu, 23 Jul 2026 18:02:06 GMT

Palantir CEO Alex Karp went on CNBC and announced a partnership with Nvidia to build an AI model that's fully air gapped for the US government. It's the biggest signal yet for a concept called sovereign AI, and it raised a question every enterprise should be asking: is my data, my prompts, and my competitive edge sitting inside someone else's model? Dhaval Kapadia and Venkatesh Sankaran recorded this emergency episode because they couldn't stop arguing about what it means.

Disclaimer: This episode was shot on July 12th, so the recent AI adoptions etc might or might not be discussed in it.

In this episode:

Why Palantir and Nvidia's government AI deal is the biggest sovereign AI signal yet
What sovereign AI actually means, explained simply
The three questions every enterprise should be asking about their data, their prompts, and their competitive edge
How Palantir is using Nvidia's open source models and air gapping them from the outside world
Will Anthropic and OpenAI ever support sovereign AI, or does it break their business model
The "fresh air" problem: how a fully closed system risks cutting itself off from innovation
Why the disagreement matters: different tiers of enterprises will adopt this differently
The realistic path for mid sized companies: on prem SLMs instead of full sovereign infrastructure
How routing only the hardest questions to the big LLMs can cut token costs by 90 percent or more
Why the hardware needed to run these models is about to get dramatically smaller
Who becomes the second, third, and fourth sovereign AI customer after the US government

Timestamps:
00:00 Why this is an emergency episode
01:24 What just happened: Alex Karp, Palantir, and Nvidia
02:53 Sovereign AI, explained simply
05:45 The three questions every enterprise should ask
08:49 How the Palantir Nvidia deal actually works
10:52 Will the big LLM labs ever support sovereign AI
11:46 The "fresh air" problem: staying open to innovation
13:43 Where the hosts disagree
16:31 The realistic path for mid sized companies
16:51 Using SLMs to cut token costs by 90 percent
19:19 Why the hardware barrier is about to disappear
20:23 What comes after the US government

Subscribe and hit the bell for weekly breakdowns of the enterprise AI news that actually matters!

The US Government Blocked the World's Smartest AI. Then Gave It Back.

Blue Skies Venture Studio — Wed, 22 Jul 2026 06:07:35 GMT

Anthropic's most powerful model was blocked by the US government within days of launch, then released again with reinforced guardrails. In episode 6 of Enterprise AI Insiders, hosts Venkatesh Sankaran and Dhaval Kapadia trace the full Fable 5 and Mythos timeline and what it means for any enterprise trying to stay at the AI frontier.

They also break down the $9 billion the biggest players have committed in a matter of weeks to place AI engineers directly inside enterprise customers, and the ROI reckoning now hitting CFOs after a year of open-ended AI spending.

In this episode:
• The Fable 5 and Mythos timeline: restricted, pulled back, and re-released
• What happens to enterprises when access to frontier models can be switched off
• Why the pause gives open-source and on-prem models room to catch up
• OpenAI's new model lineup and the same safety gating playing out again
• Forward-deployed engineers: AWS, Microsoft, OpenAI, and Anthropic committing roughly $9B in weeks
• Why this is "outsourcing in reverse" and how the cost gets baked into contracts
• The ROI reckoning: Uber's token spend, the 95% experiment failure rate, and what actually returns value
• Token cost, on-prem NVIDIA, and the cloud-vs-device calculus every CFO now faces
• Why multi-vendor strategy matters more than ever

Enterprise AI Insiders is a business show for CEOs, CXOs, and founders, not a tech show. Hosted by Venkatesh Sankaran and Dhaval Kapadia.

This episode was recorded on July 11th, 2026. Figures and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Everyone's Renting AI. The Winners Are Building It.

Blue Skies Venture Studio — Mon, 20 Jul 2026 17:18:10 GMT

Satya Nadella's viral essay, "A Frontier Without an Ecosystem Is Not Stable," argues that legacy enterprises, not AI labs, will define the next phase of AI. It's sparking a comeback for small language models and on-prem, enterprise-owned AI systems.

Databricks' Data + AI Summit doubled down with Genie, an agentic system built to learn how your enterprise actually works. And Microsoft, Salesforce, and ServiceNow are now in an all-out platform war for the enterprise AI stack, even as 70 to 80% of AI experiments are failing.

In this episode:
• Why the center of gravity is shifting from AI labs to legacy enterprises
• The "harness" concept and why AI has to live inside the enterprise
• The comeback of small language models and on-prem deployment
• The token budget problem: burning a year's spend in three months
• Databricks' Genie and the "20,000 employees, 2 million agents" framework
• The platform war between Microsoft, Salesforce, and ServiceNow
• The 5 pillars to evaluate any AI platform on: build, deploy, govern, scale, secure
• Why 70 to 80% of enterprise AI experiments are failing right now
• Global platforms beyond the West: Alibaba, Sarvam, and regional AI
• Vendor lock-in and how to manage it

Enterprise AI Insiders is a business show for CEOs, CXOs, and founders, not a tech show. Hosted by Venkatesh Sankaran and Dhaval Kapadia.

Recorded on 18th June, 2026. Figures and developments discussed were current as of that date and may have changed since. Nothing in this episode is financial, investment, or legal advice.

Anthropic and OpenAI Are Both Filing for IPOs! Here's What $100B Buys

Blue Skies Venture Studio — Thu, 16 Jul 2026 16:52:15 GMT

Claude's new Mythos and Fable 5 models can run autonomously for up to 12 hours. Apple just gave Siri a major AI overhaul at WWDC. Anthropic and OpenAI have both filed for IPOs targeting trillion-dollar valuations. Dhaval Kapadia and Venkatesh Sankaran cover the biggest enterprise AI stories of the last few weeks.

In this episode:

Claude's Mythos and Fable 5: why long-running, hours-long autonomous tasks are a game-changer for enterprise
How Fable 5 routes sensitive queries (cybersecurity, biology, chemistry) to safer models automatically
Token efficiency vs. token cost: why running longer doesn't mean spending more
Real enterprise use cases: quarterly reporting, code migration, and multi-month projects done in a day
Apple's WWDC AI overhaul: what's new with Siri and the on-device privacy angle
The Apple-Gemini partnership and what it means for enterprise data security
Why Apple's ecosystem control could be a foot in the door for enterprise adoption
Anthropic and OpenAI's IPO filings: trillion-dollar valuations and $50-100B capital raises
SpaceX's IPO and the infrastructure race behind AI
Why profitability pressure post-IPO could reshape how AI companies sell to enterprise
The coming wave of small language models (SLMs) and vertical-specific AI tools
Regulation on the horizon: the AI America Act and state-level AI laws

Subscribe and hit the bell for weekly breakdowns of the enterprise AI news that actually matters!

DISCLAIMER:

Recorded June 11, 2026. Figures and developments discussed were current as of that date and may have changed. This episode is for informational purposes only.

OpenAI Just Bought Its Way Into Consulting: Should Anthropic Be Worried?

Blue Skies Venture Studio — Thu, 16 Jul 2026 16:34:01 GMT

OpenAI is making a serious move into enterprise consulting, SAP just unveiled a 200-agent "Autonomous Enterprise," and new data shows a massive gap between AI ambition and AI readiness. Dhaval Kapadia and Venkatesh Sankaran unpack this week's three biggest enterprise AI stories.

Disclaimer: This episode was shot 1-2 months back, so the recent AI adoptions etc might or might not be discussed in it.

In this episode:

Why OpenAI acquired UK-based Tomorrowco to launch Deployco, its own forward-deployed engineering arm
Whether OpenAI's consulting push stays exclusive to its own tech, or opens the door to Anthropic and others
The data access question: what OpenAI really gains by putting engineers inside client companies
SAP's "Autonomous Enterprise": 200+ AI agents and 50 Joule assistants aiming to automate 30-70% of tasks in finance, HR, and supply chain
The Dun & Bradstreet survey finding 97% of companies have active AI initiatives, but only 5% say they're actually ready
Why the adoption gap comes down to skills, data infrastructure, and governance — not lack of investment
The rise of Chief AI Officer roles and what's next for enterprise AI in 2026

Why Big Tech Is Spending $700 Billion on AI | Enterprise AI Insiders Ep. 2

Blue Skies Venture Studio — Fri, 10 Jul 2026 22:28:45 GMT

Big Tech just committed over $700 billion to AI in a single year. In episode 2 of Enterprise AI Insiders, our hosts Venkatesh Sankaran and Dhaval Kapadia unpack the capex explosion, whether it pays off or becomes the next metaverse, and two moves quietly reshaping how enterprises adopt AI.

Venkatesh brings a 30+ year enterprise tech lens; Dhaval brings a finance and Wall Street lens. Together they cover three stories driving the AI economy right now.

In this episode:
• The hyperscaler spending surge: Amazon, Microsoft, Google, and Meta pushing past $700B, up from roughly $400B a year ago
• Where the money comes from (balance sheet, revenue, debt) and the ROI question investors can't yet answer
• Why Meta's stock fell 7% on its spending news, and whether AI is the next metaverse
• The GPU shortage, energy grid, and training demands behind the numbers
• Why Apple isn't counted among the hyperscalers
• Anthropic's $1.5B partnership with Blackstone, Apollo, Goldman Sachs, and others, and how forward-deployed engineers challenge traditional consulting firms
• How AI labs are pushing into mid-market companies through PE portfolios
• The Pentagon and Department of Commerce AI deals, why Anthropic dropped out, and why AI safety and human oversight matter from day one

The recurring theme: this shift needs more people, not fewer.

Enterprise AI Insiders is a business show for CEOs, CXOs, and founders, not a tech show. Hosted by Venkatesh Sankaran and Dhaval Kapadia. Listen and subscribe on Spotify, Apple Podcasts, and YouTube.

DISCLAIMER:

Recorded May 7, 2026. Figures and developments discussed were current as of that date and may have changed. This episode is for informational purposes only and is not financial or investment advice.

Every Company Will Become an AI Company | Enterprise AI Insiders Ep. 1

Blue Skies Venture Studio — Fri, 10 Jul 2026 22:17:37 GMT

Every company will become an AI company. In the debut episode of Enterprise AI Insiders, Venkatesh Sankaran and Dhaval Kapadia lay out why AI has stopped being a tool you plug in and has become the operating system businesses actually run on.

Drawing on Venkatesh's 30+ years building enterprise tech platforms and Dhaval's finance and Wall Street background, they argue this shift is bigger than internet, mobile, and cloud combined, and moving far faster.

In this episode:
• Why "every company becomes an AI company" is the defining shift of the next decade
• Treating AI as a tool vs. as a business operating system
• The Starbucks playbook: how a coffee chain became a tech company, and what enterprises can learn from it
• Why large enterprises, not startups, hold the real AI advantage (hint: it's the data)
• What Blue Skies AI Studio is and how it bridges startups and enterprises
• Why slow-moving enterprises risk getting left behind

This is a business show for CEOs, CXOs, and founders, not a tech show. Upcoming episodes bring guests from across the enterprise and startup ecosystem.

Listen on Spotify, Apple Podcasts, and YouTube.