The Daily AI Executive
By Stephen Adegasoye
Executive Summary
- Governance is going operational. More than 1,100 employees at OpenAI, Anthropic, Google and Meta have signed an open letter asking Washington to help build verifiable international "pacing" infrastructure for frontier AI — a sign that safety debates are moving from position papers into policy machinery your risk and compliance functions will eventually have to interface with.
- The infrastructure land-grab keeps compounding. AMD's $14bn, multi-gigawatt data-center commitment with Core Scientific is the latest evidence that AI capacity is being locked up years in advance — a dynamic that will keep upward pressure on cloud and compute line items long after per-token prices fall.
- Media economics are being rewritten in real time. Zero-click AI answers are cutting publisher ad supply sharply, even as content licensing matures into a genuine (if concentrated) revenue line — a bifurcation every rights-holder and ad-funded business needs modelled into FY27 planning now.
By the Numbers
$2.59 trillion, +47% YoY (Gartner) Worldwide AI spending forecast, 2026 | +63.4% to $64bn (Gartner) AI platforms & models market growth, 2026 | $14bn+ over 15 years, 529MW scalable to 2.5GW Core Scientific–AMD data-center deal | Down up to 40% (Ozone/Digiday benchmarking) Publisher ad supply decline, Q2 2026 |
GovernanceFrontier AI Employees Petition Washington for a "Pacing" Mechanism
What happened: An open letter circulated on 28 July 2026 and signed by more than 1,100 employees at OpenAI, Anthropic, Google and Meta asks the US government to support an international mechanism that could slow AI development in a coordinated, verifiable way if systems ever outpace safe oversight. Notably,
it is signed by more than 1,100 employees at frontier AI companies including OpenAI, Anthropic, Google, and Meta, asking the US government to support an international pacing mechanism for advanced AI development, and does not call for an immediate pause, instead asking Washington to help build the technical and governance infrastructure that would make a verifiable, coordinated slowdown possible if AI systems ever advance faster than humans can safely oversee them.
The signatories carry unusual weight:
they are not outside critics but the people building the systems.
Why it matters: This is not activist pressure from outside the industry — it is internal. When cofounders and chief scientists at rival labs co-sign a governance ask, it signals that competitive dynamics alone won't self-correct fast enough for comfort, even among believers in the technology.
Who wins: Policy-literate enterprises that get ahead of emerging oversight frameworks; governance and audit tooling vendors.
Who loses: Companies treating AI governance as a compliance afterthought; any business whose AI roadmap assumes today's permissive regulatory environment is permanent.
Commercial implications: Expect this to accelerate momentum toward model registries, usage disclosures and audit trails — all of which will eventually touch vendor contracts and procurement clauses.
Finance implications: Governance infrastructure has a cost. Budget for compliance tooling, legal review of AI vendor terms, and potential new disclosure obligations before they're mandated, not after.
Media implications: Content, likeness and IP-related AI risks (deepfakes, unauthorized training use) are likely to be early beneficiaries of any "verifiable" oversight infrastructure — relevant to any rights-holder negotiating AI licensing terms.
Long-term impact: This is the opening move in what will likely become formal international coordination mechanisms over the next 2–3 years — plan governance capability as a multi-year build, not a one-off project.
Confidence: Medium **
Sources: buildfastwithai.com AI News Today, techstartups.com Top Tech News
InfrastructureAMD Locks In $14 Billion, 2.5-Gigawatt AI Data Center Pact
What happened: AMD and Core Scientific announced a 15-year infrastructure partnership disclosed alongside Core Scientific's Q2 earnings.
Core Scientific has doubled its leased AI data center capacity to approximately 1.1 GW after signing a 15-year infrastructure agreement with AMD covering 530 MW across five campuses and representing more than $14 billion in potential base contracted revenue, announced alongside the company's second-quarter results on 28 July, with the agreement beginning in 2027 and giving AMD exclusive rights to reserve up to an additional 2 GW of future capacity.
Notably,
Core Scientific disclosed that fully delivered AI data center capacity is expected to cost approximately $11 million to $12 million per megawatt, including construction, electrical infrastructure, cooling systems, commissioning and utility integration, meaning the initial deployment represents roughly $6 billion of infrastructure investment.
Why it matters: This is the physical-capacity race made explicit — chips are no longer the bottleneck; power and shovel-ready data-center space are.
As AMD's Mathew Hein put it, "Core Scientific's extensive portfolio of AI-ready data centers expands access to infrastructure," because without guaranteed power and space, even the best chip cannot ship to customers on schedule.
Who wins: Data-center landlords and colocation operators with pre-built power capacity; AMD's ecosystem of cloud and enterprise customers seeking non-NVIDIA capacity.
Who loses: Companies without long-term power/compute contracts, who will face a tightening, higher-priced spot market for AI infrastructure.
Commercial implications: Expect continued consolidation of AI compute supply into long-duration contracts — reducing the market's flexibility and increasing switching costs for cloud buyers.
Finance implications: Multi-gigawatt, multi-year contracts of this size (structured with equity warrants, not just cash) signal that infrastructure financing is becoming as important as model licensing in AI total cost of ownership — CFOs should track capex-linked vendor exposure, not just SaaS-style opex.
Media implications: Rendering, transcoding, personalization and generative-content workloads at streaming scale will increasingly compete for the same constrained compute pool — media companies without reserved capacity may see rising unit costs for AI-driven production and localization tools.
Long-term impact: Compute scarcity — not model capability — may be the real constraint on enterprise AI ambition through 2027–2028.
Confidence: High **
Sources: Data Center Knowledge, Yahoo Finance/GuruFocus, Core Scientific investor relations
Media EconomicsAI Answer Engines Squeeze Publisher Ad Supply as Licensing Deals Multiply
What happened: Two opposing forces are reshaping content economics simultaneously. On one side,
publisher ad supply fell by up to 40% in Q2 2026 according to US and UK benchmarking data from Ozone reported by Digiday, as zero-click AI search cut the flow of traffic to news and open-web sites, with a separate study finding AI is accelerating the collapse of Google Search referrals with traffic to UK publishers projected to halve, and 37% of consumer searches now starting in AI tools instead of Google.
On the other, licensing has matured into real recurring revenue for those with negotiating leverage:
as of July 2026, OpenAI has signed roughly two dozen publisher and data deals, the single largest reportedly $250 million over five years with News Corp, while Reddit disclosed $203 million in aggregate data-licensing contract value in its IPO filing, and Amazon, Meta, Google, Microsoft and Mistral have all entered the market, with News Corp's licensing arrangement averaging around $50 million per year across its portfolio.
Why it matters: Scale determines outcome.
A market pays for scarcity and negotiating leverage, and a small publisher structurally has neither — there is one Wall Street Journal, there is one Associated Press.
For content businesses without brand-name scarcity, the licensing upside is largely closed off even as the traffic downside accelerates.
Who wins: Large rights-holders and news agencies with irreplaceable archives or real-time feeds; ad-tech platforms building "AI visibility" measurement products.
Who loses: Mid-and-long-tail publishers and ad-supported digital properties losing referral traffic with no offsetting licensing revenue.
Commercial implications: Advertising strategies built purely on open-web referral traffic need rebalancing toward owned audiences, connected TV, and direct subscription relationships.
Finance implications: Model the ad-revenue downside from declining referral traffic separately from any AI licensing upside — they accrue to different parts of the business and rarely net out for mid-sized players.
Media implications: Expect accelerating investment in "AI visibility" as a new brand and reach metric, alongside continued programmatic expansion into connected TV inventory as a hedge against open-web decline.
Long-term impact: The content economy is bifurcating into a small tier of well-compensated scarce IP holders and a much larger tier absorbing the traffic cost with little compensating revenue — commercial finance teams should stress-test five-year ad revenue models against this split now.
Confidence: Medium **
Sources: Digiday (via ProOps Consulting synthesis), LLM Pulse licensing tracker, Press Gazette
Deep Dive: From Flat-Fee AI to Metered Compute — Why Your 2027 AI Budget Line Will Look Nothing Like Your 2026 One
For two years, enterprise AI felt like software: a flat monthly fee per seat, roughly predictable, easy to forecast. That era is ending, and every commercial finance leader needs to understand why.
First principles. Model providers are caught between two forces pulling in opposite directions:
1. Per-token prices are collapsing.
AI inference costs have fallen roughly 95% in two years and about 1,000x in three years for a fixed capability level, from $30 per million tokens for GPT-4 in 2023 to under $0.50 for equal-quality open models in 2026.
2. Total enterprise spend keeps rising anyway. This is because usage volume, not unit price, now drives the bill. Agentic workflows that reason for longer and call tools repeatedly consume vastly more tokens per task than a single chat message did.
That combination — falling unit price, rising total consumption — is why finance teams are simultaneously being told prices are dropping and watching their AI invoices grow.
A growing number of enterprise customers are cutting spending on OpenAI and Anthropic, switching to cheaper AI alternatives and demanding clearer returns on investment; one AI startup CEO moved 100% of his company's traffic from Anthropic's Claude models to a cheaper Chinese alternative, saving millions of dollars within months, calling it "a matter of survival for the business."
Meanwhile
Uber's CTO burned through the company's entire 2026 AI budget by April.
The billing model is shifting underneath you.
Anthropic has shifted some business customers toward actual-usage billing, GitHub has introduced a new usage-based system after monthly allotments, and OpenAI, Anthropic and Microsoft's GitHub have started moving beyond simple flat-fee subscriptions, with heavier users potentially paying more when AI tools produce slide decks, draft emails, debug code or run longer agent-based tasks.
Gartner's own research confirms the discipline this is forcing on buyers:
"Enterprise AI budgets are coming under greater scrutiny, with increased focus on usage efficiency, cost control and measurable outcomes," with spending shifting toward providers who can demonstrate clear value across cost, latency, performance and reliability.
A simple framework for the transition:
| Old model (2023–2025) | New model (2026 onward) |
|---|
| Flat per-seat subscription | Usage-metered, credit-based billing | | Budget = headcount × fixed fee | Budget = workload volume × variable rate | | Vendor lock-in via habit | Vendor switching via price/quality arbitrage | | Procurement owns the decision once a year | FP&A must monitor consumption monthly | | AI cost treated like software licensing | AI cost treated like a utility (compute, power, water) |
Why executives should care. Compute now represents the dominant cost of running an AI company —
in both China and the US, compute makes up more than 50% of the cost of running an AI company, with R&D and inference compute together accounting for 54–62% of total operating costs across three AI companies where estimates were possible.
Vendors facing that structural cost base cannot subsidise flat-fee plans forever. Expect continued repricing, tiering, and metered billing to spread across every major provider through 2027.
Commercial Finance Implications
Three opportunities:
1. Renegotiate for usage transparency, not just price. As billing shifts to metered models, insist on granular usage dashboards and cost-per-workflow reporting in every renewal — this is the single highest-leverage FP&A ask right now.
2. Arbitrage the commoditization. With inference costs having fallen dramatically for comparable capability, route lower-stakes tasks (transcription, first-draft copy, localization) to cheaper open-weight or mid-tier models and reserve premium frontier models for high-value, high-risk work.
3. Monetize scarce IP directly. If your organisation holds distinctive archives, live feeds, or brand-name content, the licensing market — while concentrated — is real and growing; a direct commercial conversation with AI labs may outperform waiting for ad-market recovery.
Three risks:
1. Budget-blind agentic rollouts. Agent-based workflows can consume tokens unpredictably; without hard usage caps, a single automated process can exceed a quarter's allocated AI spend, as happened at Uber.
2. Referral traffic erosion with no offsetting revenue. Mid-sized, ad-funded content businesses risk losing meaningful traffic to zero-click AI answers without qualifying for licensing deals that only reward scarcity.
3. Vendor concentration risk disguised as convenience. Locking into a single provider's flat-fee plan now may look cheap, but metered repricing could arrive with little warning once promotional pricing ends.
Three ideas to explore:
1. Build a rolling "AI cost-per-output" dashboard (cost per finished asset, per localized minute of content, per generated ad variant) rather than tracking vendor invoices in isolation.
2. Pilot a model-routing layer that automatically directs tasks to the cheapest model meeting a defined quality bar, with human review gates preserved for brand-sensitive or IP-related outputs.
3. Establish a standing quarterly review between finance, legal and content/rights teams to reassess AI licensing and vendor terms as this market remains structurally in flux.
Executive Talking Points
1. AI infrastructure spend is now a capex conversation, not just an opex one — multi-gigawatt, multi-year contracts mean compute commitments increasingly resemble real estate or power-purchase agreements.
2. Falling per-token prices are not the same as falling total AI spend — plan for both trends simultaneously.
3. Governance is moving from PR statement to policy infrastructure — treat it as a compliance roadmap item, not a talking point.
4. Content and rights-holders face a bifurcating market: scarce, brand-name IP is being compensated; everything else is absorbing traffic loss with no offset.
5. Vendor lock-in is evaporating as models commoditize — procurement leverage has shifted meaningfully toward buyers for the first time in this cycle.
AI Tool of the Day
ChatGPT Work for Finance teams (OpenAI)
What it does: A finance-specific ChatGPT Enterprise configuration built for forecasting, month-end close support, variance analysis and board-ready reporting, with source-grounded outputs and human review checkpoints.
Who it's for: FP&A teams, controllership, and finance business partners.
Pricing: Bundled within ChatGPT Enterprise/Business tiers; specific finance-feature pricing not publicly disclosed at time of writing.
Why it matters:
Time saved per message varies by function, with accounting and finance users reporting the largest benefits, and on average ChatGPT Enterprise users attribute 40–60 minutes of time saved per active day to their use of AI.
Should a finance leader learn it: Yes — directly applicable to close cycles and forecasting workflows.
Time required: 2–3 hours for a working pilot on a single recurring report.
ROI: High for repetitive reporting and variance-explanation tasks; requires human validation for any external-facing or audited output.
AI Paper / Report of the Day
Source: Epoch AI, "LLM inference prices have fallen rapidly but unequally across tasks" plus associated 2026 data insights on hyperscaler capex and AI company cost structures.
Problem: Understanding whether falling per-token AI prices genuinely reduce enterprise cost exposure, or merely mask a different, rising cost base.
Method: Epoch AI tracked state-of-the-art model performance against six benchmarks over three years, measuring the price required to reach fixed capability levels, alongside separate tracking of hyperscaler capital expenditure and AI company operating cost structures.
Findings:
Hyperscaler capex has quadrupled since GPT-4's release, nearing half a trillion dollars in 2025, with the combined capital expenditures of Alphabet, Amazon, Meta, Microsoft, and Oracle growing at an average of 72% per year — a trend that, if it continues, would mean they collectively spend $770 billion in 2026.
Separately,
hyperscaler capex is on trend to outpace their cash inflows by the end of 2026.
Why executives should care: This is the clearest evidence available that the AI cost curve for buyers (falling per-token prices) and the AI cost curve for builders (rising capex, thin margins) are diverging sharply — a structural tension that will keep resurfacing as pricing wars, usage-based billing and vendor consolidation.
Build Something
Exercise: Build a one-page "AI cost-per-output" tracker (25 minutes)
Take your organisation's current AI vendor invoices and map spend against actual business outputs — cost per localized minute of content, cost per generated ad variant, cost per automated report — rather than simply tracking total monthly spend by vendor. This reframes AI cost conversations around unit economics rather than headline invoice size, which is the language your board and your vendors are increasingly using.
Time required: 25 minutes to build a first draft in a spreadsheet using last month's actual usage data.
Why it matters: Once AI billing shifts fully to usage-based models, cost-per-output will be the only metric that lets you compare vendors, models and internal workflows on equal terms.
Skill of the Day
Model routing
Why: As pricing and capability converge across providers, the highest-value skill for both technical and finance teams is knowing which tasks need a premium frontier model and which can run on a cheaper, faster alternative without loss of quality.
Difficulty: Medium — requires basic understanding of model capability tiers and workflow classification, not coding expertise.
Time to learn: 3–4 hours for foundational literacy; ongoing refinement as new models launch.
Best resource: Vendor rate-card documentation (e.g., OpenAI and Anthropic's published model tiers) combined with internal usage-pattern analysis from your own AI cost dashboard.
Executive Quote
"Enterprise AI budgets are coming under greater scrutiny, with increased focus on usage efficiency, cost control and measurable outcomes." — Arunasree Cheparthi, Senior Principal Research Analyst, Gartner
Sources
What You Should Do Today
1. (15 min) Pull your last three AI vendor invoices and check whether billing is still flat-fee or has begun shifting to usage-based credits — flag any upcoming renewal dates for renegotiation.
2. (20 min) Ask your content or marketing team whether any AI licensing conversations are underway with your archives or proprietary data, and whether legal has reviewed the terms against the News Corp/Reddit benchmarks now public.
3. (25 min) Draft the first version of a cost-per-output tracker (see Build Something) using one AI-enabled workflow already in production, to establish a baseline before usage-based billing fully arrives.
|