Gilles

Tech exec | CTO @ ADT | Ex-Google, Zendesk, Motorola, Apple, Adobe | Building platforms at the edge of AI, security, and the connected home | SF-based

[Gilles @gilles]

I'm excited to speak at #Unify2026, the inaugural event from the Connectivity Standards Alliance bringing together the companies, technologies, and conversations shaping the future of the IoT. From interoperability and open standards, to AI and connected ecosystems, our industries have a real opportunity to come together and create trusted, reliable outcomes for our customers in life's most critical moments.

Register today and join me June 16-18 in Austin!

https://t.co/uLsLTQK87k

#Unify2026 #csaiot #standardsmatter

[Sam Altman @sama]

Here is our current plan for OpenAI:

https://t.co/r29FUUee3A

[Gilles @gilles]

Reports suggest Siri may reportedly be powered by Gemini, and if that rumored arrangement proved real, it would hand the inference layer to Google right where the user relationship lives. Apple would own the device and the customer. Google would own the model. When something goes wrong, accountability would sit across an org boundary no keynote will resolve. Apple has engineered out of supplier dependencies before, but that work takes time, and the exposure would be real in the interim.

Read more: https://t.co/2KbdkfAqi0

#AppleIntelligence #PlatformStrategy #FrontierAI #EnterpriseAI #WWDC

[Gilles @gilles]

This is like early cloud, except the scarcity is structural, not temporary.

In cloud, demand was predictable enough to pre-build capacity. AI inference demand is outrunning permitting, power, and chip delivery cycles simultaneously. Even $180B in annual capex doesn't close that gap fast enough.

So Google becomes a customer of a direct competitor. Anthropic signs $45B with the same provider. These aren't vendor contracts, they're entangled platform dependencies formed under supply pressure.

The structural question: whoever controls GPU inventory and power at this scale has leverage over the pace of AI deployment across the whole industry. Not by design. By arithmetic.

Read more: https://t.co/gtkpMfEeOR

#AIInfrastructure #ComputeStrategy #EnterpriseAI #GPUMarket #PlatformStrategy

[Gilles @gilles]

Broadcom guarantees the shortfall if chip liquidation falls short, but only for A1 and A2 note holders. The $4.5B B tranche sits outside that protection entirely.

The $35B gets headlines. What matters: lease-backed repayment for senior investors, Broadcom absorbs their liquidation risk, Atlas SP Partners puts in $800M equity to own the SPV, and Anthropic gets compute capacity without contributing its own equity.

The stress scenario: chip cycles outrunning the amortization schedule means the guarantee gets called in ways the original model didn't price. At this vendor concentration, any lease renegotiation ripples through the whole capital stack.

The broader shift is real. This looks like project finance now, asset-backed tranching, guaranteed floors, long-dated leases doing work that equity used to carry alone.

Read more: https://t.co/FyU8nJtNGI

#AIInfrastructure #PrivateCredit #FrontierAI #Semiconductors #CapitalMarkets

[Gilles @gilles]

A headline at The Hacker News references CVE-2026-20245 in Cisco Catalyst SD-WAN Manager as actively exploited with no patch available, but the source page contains no article body and no supporting evidence for those claims. The headline alone is not sufficient to confirm them. Check Cisco's official advisories for verified information.

Read more: https://t.co/QUDDxmcWZt

#Cybersecurity #NetworkSecurity #EnterpriseArchitecture #SDWAN #ZeroDay

[Gilles @gilles]

Google is contracting roughly $920 million a month in Nvidia GPU compute from SpaceX through mid-2029, per Bloomberg citing SEC filings. SpaceX built out substantial GPU clusters for its own AI and Starlink work and is now supplying that capacity to Google for AI training and inference. When the most sophisticated infrastructure operator on the planet starts signing structured, multi-year compute dependencies on a third party, the hyperscaler is no longer the terminal node. It is a middle layer. Enterprise AI roadmaps are probably underestimating how unsettled the supply layer still is.

Read more: https://t.co/kaAf8i6hdi
#CloudInfrastructure #AIInfrastructure #EnterpriseAI #PlatformStrategy #Semiconductors

[Gilles @gilles]

"The first major vendor to normalize always-on agents in the enterprise sets the terms for everyone that follows."

That is the actual story. Not the leaked "make people addicted" framing.

Microsoft's reversal on OpenClaw is competitive exposure calculus, not changed beliefs. Scout's Entra identity layer plus Work IQ compounding over time is lock-in architecture, not a product feature.

Most enterprise teams are underestimating how quickly a two-year-deep semantic model of your org becomes structurally irreplaceable.

Read more: https://t.co/T4Y3Ovw6eD

#EnterpriseAI #AgenticAI #PlatformStrategy #Microsoft #EnterpriseArchitecture

[Gilles @gilles]

Nadella's Build interview is worth reading carefully. The MAI model strategy isn't about benchmark competition. It's about owning the layer where enterprises train their own models on proprietary data over time.

Stage 1: model access. Stage 2: workflow integration. Stage 3: continuous learning loops on your own data. Most enterprises are still between 1 and 2.

The real bottleneck isn't the tooling. It's that most organizations don't have the data discipline or ML ops maturity to actually run these RL loops at scale.

Switching costs that accumulate at the model-training layer are structurally deeper than switching costs at the application layer. That's the architecture play here.

Read more: https://t.co/9RFxSobhEV

#EnterpriseAI #PlatformStrategy #Microsoft #AIInfrastructure #MLOps

[Gilles @gilles]

The agentic AI security problem isn't the model. It's the infrastructure the model runs inside.

Enterprise controls depended on human hesitation as an accidental safety layer. Nobody designed it that way. Agents remove it without anyone noticing.

What this means in practice:

• Mosaic effect: agents assembling authorized data points into conclusions no single permission intended. Object-level audits miss this entirely.
• Read/act split must be architectural. Summarizing a doc and transmitting it are different actions. Most systems don't enforce the difference.
• Rollback isn't remediation. Behavioral residue persists in memory after a system prompt reset.

37% of orgs have agents deployed. 3% have agent-specific controls. That gap is where the Monday morning incident is waiting.

Read more: https://t.co/mdzM5rfalF

#AgenticAI #EnterpriseSecurity #AIGovernance #CyberSecurity #EnterpriseAI

[Gilles @gilles]

276,000 people. One AI platform. Embedded in the client workflow layer, not handed out as a chatbot.

The structure is what matters: Claude inside Digital Gateway, where client data and billable work already live. That's a platform dependency decision, not a productivity one.

The PE angle makes it concrete. KPMG Blaze uses Claude Code to modernize legacy IT in portfolio companies. A services firm making its AI stack the delivery mechanism for transformation work it used to staff.

The hard part isn't building the agent. It's governing it across 138 countries with client data sensitivity varying by jurisdiction. Most enterprise planning cycles are underestimating that friction.

Read more: https://t.co/nEw5EYLefg

#EnterpriseAI #AIStrategy #Anthropic #ProfessionalServices #AIDeployment

[Gilles @gilles]

Alphabet raised $85B in external equity despite sitting on $127B in cash. The real signal isn't the capital raise.

The visible fact: capex goes from $31B in 2022 to $185B in 2026. Supply is behind demand.

The deeper driver: on-premises TPU deployment, a $462B cloud backlog, deal velocity doubling in the $100M-$1B range. This is platform lock-in at infrastructure scale, not capacity expansion.

Token volumes went from 9.7 trillion/month two years ago to 3.2 quadrillion today. Serving costs fell 78% in 2025. The unit economics now support scaling to surfaces that weren't viable before.

Most enterprise architecture decisions made in the next 18 months will anchor to one of two or three infrastructure relationships for the better part of a decade.

Read more: https://t.co/u13zzDlctL

#EnterpriseAI #CloudInfrastructure #PlatformStrategy #AIInfrastructure #GoogleCloud

[Gilles @gilles]

"The time has come for every company to move from consuming a frontier model to fully participating at the frontier." , Nadella at Build 2026.

The surface read: Microsoft competing with OpenAI and Anthropic. The structural read: Microsoft hedging its dependency position before both IPO and reprice.

Azure is becoming a routing layer between Microsoft-owned and third-party models. That's not a model selection decision for enterprise teams. It's a platform dependency decision.

Most enterprise planning cycles are probably still treating it like the former.

Read more: https://t.co/IlCvf7sNCX

#EnterpriseAI #PlatformStrategy #FrontierAI #AIInfrastructure #AzureAI

[Gilles @gilles]

When multiple agents query the same database and return different answers, who owns the definition of what the data actually means?

That is the production failure mode most enterprise AI teams are not tracking yet. Agents confidently returning different answers from the same schema is a governance failure, not a model failure.

Every major vendor is now racing to own the context layer while claiming openness. Snowflake anchors it in the catalog. Microsoft in Fabric. The lock-in surface is shifting from the model to the semantic layer.

Most deployments running multiple agents against shared data are silently accumulating semantic drift today.

Read more: https://t.co/LQ0W3jfGTm

#EnterpriseAI #AgenticAI #DataGovernance #PlatformStrategy #AIDeployment

[Gilles @gilles]

Enterprise AI right now looks like enterprise ERP in the late 1990s, except the failure mode is reversed.

With ERP, the technology was too rigid for the org. With AI, the org is too rigid to absorb what the technology produces.

$252B invested in 2024. Only 6% of firms with significant EBIT impact. Project abandonment jumping from 17% to 42% in one year.

Learning investment was associated with 34% higher AI benefit likelihood, nearly twice the 19% lift from more infrastructure spend. Most budgets are inverted relative to where the constraint actually lives.

Read more: https://t.co/NmOGYQk9nW

#EnterpriseAI #AIStrategy #Leadership #OrganizationalDesign #DigitalTransformation

[Gilles @gilles]

Two years. A remote code execution flaw sat undetected in Redis, one of the most scrutinized open-source components in enterprise infrastructure, before an autonomous AI tool found it.

The number that matters is not the CVE severity score. It is the latency gap.

"No known vulnerabilities" is a materially weaker assurance than it was 18 months ago if the audit methodology that produced it was running at human speed.

Existing CVE backlogs are probably underestimated. Not because new flaws were introduced, but because the search space AI accesses is qualitatively different from what human review covers.

Read more: https://t.co/LLCPmjEEED

#Cybersecurity #AISecOps #VulnerabilityManagement #EnterpriseAI #RiskManagement

[Gilles @gilles]

Two AI measurement frameworks: cost-per-token measures infrastructure spend. Tokens-per-result measures business outcomes.

Most enterprise AI budgets are still optimized around the first one.

Labs optimizing for benchmarks may be building toward a metric that enterprise buyers are actively moving away from as inference costs commoditize.

The useful dashboard tracks cost-per-completed-task, not cost-per-token. That is not a harder metric to build. It is just a harder conversation to have with the team that owns the GPU allocation.

Read more: https://t.co/g6CwSAM4vK

#EnterpriseAI #AIStrategy #Infrastructure #AIDeployment #PlatformStrategy

[Gilles @gilles]

The Microsoft 365 debug flag story isn't really about one shipping mistake.

It's about a structural gap in how enterprise security teams apply vendor trust:

• First-party apps from tier-1 vendors get lighter scrutiny than third-party apps. That asymmetry is a policy gap, not just a governance oversight.
• MDM and conditional access don't cover token-level flaws inside the apps themselves. Different layers, different exposure.
• Platform consolidation reduces complexity and concentrates blast radius at the same time. That's not a tradeoff that disappears after the patch.

Read more: https://t.co/N6JxqqLIVo

#EnterpriseSecurity #ZeroTrust #MobileSecurity #CyberRisk

[Gilles @gilles]

"We were allowed to train models at a larger scale and explicitly pursue superintelligence entirely with our own IP, with our own data, no distillation, training from scratch." , Mustafa Suleiman

That is not a product quote. It is a dependency restructuring announcement.

The real question for enterprise buyers is not which model wins the benchmark this week. It is which platform owns the workflow layer when commitments get locked in.

Most organizations are probably underestimating how fast the dependency structures forming right now will constrain their optionality in 18 months.

Read more: https://t.co/UDd1uICzkH

#EnterpriseAI #PlatformStrategy #MicrosoftBuild #AIInfrastructure #FrontierAI

[Gilles @gilles]

Meta monetizing agents inside WhatsApp and Messenger looks like a distribution deal. It is closer to a dependency structure.

Businesses routing customer conversations through Meta Business Agent are handing over intent signals, behavioral data, and relationship history to a platform they do not control.

The agent works. The switching cost in year three is the real question.

Most enterprise teams are probably underestimating how fast a distribution partnership becomes a hard dependency.

Read more: https://t.co/eCDi8A1AhW

#EnterpriseAI #PlatformStrategy #AIAgents #FrontierAI