GPT-5.5PriceHikeandAgentTollsBreaktheAIROIMath

◆ DEEP DIVES

1. 01

   The Three-Layer AI Cost Crisis — Your Gross Margins Are About to Get Audited

   The Cost Stack Just Became Three Landlords Deep

   The enterprise AI cost thesis broke this week, and it broke from three directions at once. The mechanism matters, because the portfolio response differs by layer.

   Layer 1: model provider costs are rising, not falling. GPT-5.5 ships a headline two-times price increase. The honest number, once completion-token effects are included, is 49-92% net cost inflation. OpenAI is explicitly pivoting from market share to margin. Anthropic killed flat-rate subscriptions the same week. GitHub is rebuilding its platform for 30x more load, which is what metered pricing produces at scale. A twenty-dollar Claude subscription burning hundreds to thousands in compute was never going to last. The subsidy ended.

   Layer 2: the enterprise incumbents are imposing per-action tolls. ServiceNow's Action Fabric meters every action an external AI agent takes against its data. SAP now bans external agent access without endorsement. Workday and HubSpot are both moving to usage-based agent pricing. DataDog caps MCP requests at 5,000 daily. JPMorgan's Mark Murphy named it correctly: a tax on customers using outside AI agents.

   Layer 3: the actual ROI does not cover the bill. KKR's Pete Stavros told Milken the AI earnings uplift across the portfolio is 5%, not 50%. That is the first honest number from a tier-one GP with every incentive to round up. Uber's CTO admitted they 'blew through' the AI budget after turning on agentic tools. Sierra's Bret Taylor called the ramp 'pricey before ROI.'

   ---

   What This Means for Your Book

   Every AI application company now faces a cost stack with two independent landlords who can raise rent unilaterally — the model provider and the data incumbent — selling into a buyer measuring single-digit ROI against double-digit cost increases. The math:

   A company paying 49-92% more for inference, plus new per-action fees from ServiceNow or SAP, selling into a buyer measuring 5% uplift, is running negative gross margin on AI features that were supposed to defend ARR.

   The a16z speedrun cohort confirms it at seed: founders are spending $300K/year on agents in lieu of engineers, with Cursor bills exceeding engineer salaries. Token spend and headcount are both rising at once. The 'AI replaces headcount' thesis is empirically dead at the company level while platform costs compound.

   Where Sources Diverge

   There is a reasonable counter-thesis, and it is worth stating cleanly before dismissing it. Inference costs fall fast enough that the flat-rate model looks retroactively sensible by late 2026. This is the $700B hyperscaler capex bet, that raw compute supply bends the curve. Several sources cite it as possible. None underwrite it as base case. The more honest framing, or rather the less flattering one, is that usage-based pricing sticks, wrappers pass costs through, and a meaningful cohort of AI-feature companies discover the feature that defended ARR is the feature that compressed the margin they were defending ARR to protect.

   Action items

   • Run a pricing-model stress test across every AI-exposed portfolio company this week — model gross margin under metered pricing vs. current flat-rate, with 20-40% platform toll stack added
   • Mandate multi-model routing architecture for any portco with >30% API COGS from a single provider within 90 days
   • Re-underwrite every active LBO model to ≤5% AI EBITDA lift over 24 months; present adjusted returns at next IC
   • Source 3-5 MCP gateway / agent-governance startups for active diligence at pre-A to Series A

   Sources:Benedict Evans · Laura Bratton · Cory Weinberg · TLDR AI · a16z speedrun · StrictlyVC

2. 02

   Palantir 85% vs. SaaS -30%: The Market Is Pricing a Permanent Split

   One Quarter Does Not Make a Category, But It Might End One

   Palantir printed eighty-five percent revenue growth and eight hundred ninety-two million dollars in quarterly free cash flow, a single-quarter FCF figure larger than all of Q1 2025 revenue. Management walked full-year guidance from sixty to seventy-one percent and guided Q2 to seventy-nine. US commercial grew a hundred and thirty-three. US government grew eighty-four. The stock fell after-hours anyway, because at forty-three times forward sales, roughly double the better-positioned cybersecurity comps, the company has to keep accelerating to hold the line it is already standing on.

   Meanwhile the enterprise software index is down more than thirty percent year to date. ServiceNow is down forty with an unproven agent pivot and a twenty-to-thirty percent price hike. The rest of the cohort growing at twenty percent is being marked not as cyclically weak but as structurally impaired. Karp's on-call jab, 'if you want old-school software that actually doesn't work and probably will disappear, there are a lot of names,' is aggressive enough to look embarrassing if the tape turns.

   ---

   The Bifurcation Thesis

   Multiple sources converge on the same read, or rather the more interesting version of it: the market has stopped treating enterprise software as a valuation debate and started treating it as a category mortality question. The mechanism is specific. Palantir's data orchestration model, with forward-deployed engineers, government clearances, and a proprietary data position, compounds with AI. The median SaaS incumbent's seat-license model does the opposite, because AI agents eliminate seats.

   Metric	Palantir	SaaS Peers	Cyber Leaders
   Revenue growth	85%	~20%	20-30%
   FCF quality	$892M/qtr	Compressing	Strong
   Forward sales multiple	43x	4-8x (compressing)	~20x
   AI positioning	Category winner	Defensive/disrupted	Neutral benefactor

   The less comfortable reading, and the one worth holding as a hedge: Palantir at forty-three times forward sales with management taking a victory lap is the classic top-of-cycle setup. The fundamentals are real. The multiple is not durable if growth decelerates even modestly. And seventy-one percent full-year guidance is a bar that compounds quarter on quarter.

   The Karp Line Worth Noting

   Karp named OpenAI's Deployment Company as a direct competitor on the earnings call, the first public acknowledgement that frontier labs are moving onto Palantir's enterprise deployment turf. That changes who writes the SOW. Two better-capitalized entrants with superior models just copied the GTM. Government is fine. Commercial renewals are the exposed surface.

   The market is telling you legacy SaaS growing below thirty percent without a credible AI-native roadmap belongs at four to six times forward revenue, not ten to twelve. That is not a forecast. It is a description of where the flows are already going.

   Action items

   • Reprice every SaaS position in the book: sub-30% growers get marked to distressed multiples this quarter; benchmark AI-native platforms to Palantir's growth/FCF profile
   • Screen for the 'second Palantir' in sanctioned enterprise data access — SAP-native, Salesforce-native integration layers with proprietary data position
   • Trim Palantir public exposure or buy protection above 43x forward sales; the risk/reward is asymmetric to the downside despite fundamentals

   Sources:Martin Peers · The Information AM · Morning Brew · Cory Weinberg · Mindstream

3. 03

   OpenAI-Microsoft Terms Are Public — The Cloud AI Chessboard Moved

   What the Restructuring Actually Says

   The Microsoft-OpenAI renegotiation is now specific enough to underwrite against, which is the first useful thing anyone has been able to say about it in months. The terms: Azure exclusivity is replaced by a nonexclusive IP license through 2032, OpenAI can serve any cloud including AWS, Microsoft retains roughly 27% equity in the for-profit entity and keeps "primary cloud" status, the AGI-triggered kill switch is gone, Microsoft stops collecting revenue share while OpenAI continues paying one through 2030, and the fifty-billion-dollar OpenAI-Amazon deal is now unblocked. That last item is the one that actually moves.

   This is not incremental news. It is a structural change in the cloud AI competitive landscape, and any thesis underwriting Azure lock-in was quietly repriced the afternoon the ink dried. AWS responded inside the same news cycle by distributing ChatGPT APIs on Bedrock alongside its Frontier agent tool and Stateful Runtime Environment. The optionality Microsoft gave up is now AWS inventory.

   ---

   The Foundation Model Price War Escalates

   In the same window, xAI launched Grok 4.3 at $1.25/$2.50 per million tokens, roughly fifty percent below Anthropic's Sonnet 4.6 with claimed performance parity, a 1M context window, and reasoning. OpenAI shipped an import feature in Codex that pulls settings, plugins, agents, and project configuration directly out of Claude Cowork, which collapses switching costs to approximately zero. Anthropic is being squeezed on price at the API layer and on lock-in at the workspace layer in the same news cycle.

   Meanwhile Opus 4.7 shipped with 43% more user frustration than its predecessor on Base44's experiential benchmark. Quality perception is degrading at the worst possible moment for a round.

   The Uncomfortable Valuation Math

   OpenAI carries an $833-852B implied valuation. Anthropic is targeting $900B. Both numbers price monopoly-style capture of enterprise revenue through the PE channels discussed last week. The moat they imply is being filled from below by DeepSeek V4 and open-weights frontier parity, and from above by regulation and pre-release vetting. The labs are raising at peak-cycle marks into peak-cycle moat compression. This is probably wrong, but it is the shape of the trade.

   The trade is to rebalance toward application-layer data moats, AWS-native infra, and ARM inference before the IPO comps arrive. Secondaries at meaningful discounts to primary marks; avoid primary at these valuations.

   Meta's multi-million-unit AWS Graviton ARM CPU deal for inference is the first credible dent in Nvidia's inference dominance at hyperscaler scale. It also signals the training-to-inference pivot that converts AWS's architectural bet from liability to moat. That is where the capital should go.

   Action items

   • Audit portfolio exposure to Azure-OpenAI exclusivity narratives and re-score any company whose moat was 'Azure-OpenAI co-sell partner' — present findings at next partner meeting
   • Re-underwrite Anthropic secondary positions with 20-30% Claude Cowork user churn scenario and Grok-tier pricing as baseline gross margin assumption
   • Accelerate diligence on AWS Bedrock-native agentic tooling and ARM/Graviton-optimized inference plays at Series A/B

   Sources:Last Week in AI · ben's bites · Benedict Evans · Ben Thompson · TLDR AI · Simplifying AI

4. 04

   Amazon's Inference Moat + Logistics Platform: The Re-Rating Case

   Three Amazon Signals Converging Into One Thesis

   Amazon produced three signals this week that the street is pricing as three separate stories and which are, or rather are more usefully read as, a single one: convert marginal costs into capital costs, then rent the capex. First, ASCS formally productizes logistics-as-a-service with P&G and 3M as anchor customers, and FedEx and UPS traded accordingly, with UPS closing -10.47% in a single session. Second, the AI compute market shifted from training-dominant to inference-dominant, at which point AWS's disaggregated architecture (Nitro plus custom Trainium silicon), until recently described as a weakness, starts looking like a structural cost moat. Third, Amazon Leo's 50-plus planned launches quietly supply drone-delivery connectivity.

   The unifying read: this is the third run of the retail-to-FBA-to-AWS platform playbook. If pattern recognition is alpha, ASCS is it.

   ---

   The Cloud Neutrality Trade

   The underappreciated reframing is compute neutrality as moat. Microsoft missed Azure growth earlier in 2026 because it diverted compute to internal OpenAI workloads, which is a polite way of saying the AI factory and the neutral cloud are the same building. Google has the same problem dressed up as search defense. Amazon, whose core business is moving physical goods, has no comparable internal compute sink competing with customer workloads.

   Hyperscaler	AI Compute Conflict	Inference Position	Neutrality
   AWS	Minimal	Architecturally favored	High
   Azure	High (OpenAI crowds out)	Training-optimized	Low
   GCP	Medium (Search + Gemini)	TPU (captive)	Medium

   The logistics unbundling is less glamorous and probably the larger repricing. Amazon entering the $1.3 trillion third-party logistics TAM as a formal platform participant, rather than the accidental one it has been for years, turns FedEx and UPS into structurally impaired acquirers. Any portfolio company holding a parcel-logistics exit thesis needs that thesis rewritten this quarter.

   Thesis-Kill Risk

   This is probably wrong, but the entire AWS inference-moat argument rests on one line item: Trainium 3 trajectory. If the custom silicon stalls, AWS reverts to being an Nvidia reseller with worse training economics than Oracle, which is a sentence nobody wanted to type. Track it quarterly with a hard kill threshold. The secondary risk is simpler: Jensen Huang's "won't make that mistake again" line means Nvidia intends to compete with its own customers for model-layer value.

   Action items

   • Reweight public cloud exposure: increase AMZN, reduce MSFT on Azure-neutrality risk; model 100-200bps AWS margin expansion from inference mix shift
   • Stress-test logistics and 3PL-adjacent portfolio companies for Amazon ASCS displacement over 2-4 quarter horizon
   • Build a Trainium generation-over-generation benchmark tracker as the single-point-of-failure monitor for the AWS inference thesis

   Sources:Ben Thompson · TLDR · Bloomberg Technology · Morning Brew