The Agentic Enterprise AK · Morning Edition · 7 min read | Wednesday, July 22, 2026 The build-versus-buy math is starting to favor build. Starbucks is still writing its own AI software to chase a $400 million bill, and this week the build column got cheaper still: open-weight models arrived at serving cost, and Google cut its workhorse Gemini price again. The build-versus-buy math that almost always said buy is being rerun across the Fortune 500, and every few days another input moves. Databricks is raising at $188 billion to sell the build layer. Kimi K3, DeepSeek V4, and Inkling just put frontier-class capability on servers you own. Even the buy side is repositioning: Microsoft committed billions to Mistral this week to sell control, not raw capability. The catch is unchanged. For most companies, building is still a bill, not a discount. | | The Big StoryDeals / Strategy |
The enterprise just remembered it can build. | F | or twenty years the safe answer to "build or buy" was buy. Licensing beat maintaining. That assumption is now under review at the top of the Fortune 500. Starbucks is developing its own AI software to replace a Microsoft inventory system and an IBM maintenance tool, aiming at a $400 million annual software spend, first reported by Bloomberg on July 9. Software stocks dipped on the read-through. The company frames it as cost discipline. The market read it as a warning shot at the per-seat licensing model. |
What changed is the cost of building, and this week it changed again. Databricks signed a term sheet to raise at a $188 billion valuation, funding the exact toolkit for building internal software without a vendor contract: Unity AI Gateway for governance, Genie for turning data into action, Lakebase for agents. Then the open-weight labs moved as a pack. Moonshot's Kimi K3 went live at 2.8 trillion parameters, with weights promised July 27; DeepSeek's V4 ships MIT-licensed weights; Thinking Machines released Inkling under Apache 2.0. That is frontier-class capability at serving cost, self-hosted. Google, meanwhile, cut its Gemini Flash workhorse price again. When the build tools are this capitalized, this good, and now this cheap, the license you renew every year looks like a choice, not a necessity. The counterweight is on the buy side, and it is not standing still. Uber burned its full 2026 AI budget in four months, then capped employees at $1,500 a month per tool, a reminder that building carries a metered bill too. And the vendors are repositioning: Microsoft committed billions to Mistral this week to offer European enterprises frontier AI they can run on local, air-gapped servers, selling control rather than raw capability. So the honest calculus for a CIO is not "build is free." It is that both columns now carry real, metered, board-visible cost, which is exactly why the comparison is being run again. When the tools to build get cheap, the license you renew every year becomes a decision, not a default. |
The Spearhead Take Build-versus-buy did not flip because building got easy. It flipped because buying stopped being predictable, and this week the build side got cheaper still. In the systems we ship, the win is rarely a from-scratch rebuild; it is owning the thin layer of workflow and data that is specific to you and renting the rest. Starbucks is not writing its own operating system. It is reclaiming the parts of its stack where a generic SaaS tool was overcharging for undifferentiated work. Open weights and cheaper inference just widened the set of parts worth reclaiming. That is the question worth putting on your renewal list this quarter. |
| The Obvious & The Overlooked Three reads the market has. Four it is missing. The Obvious Starbucks wants to cut a $400M software bill by building. The target is Microsoft and IBM systems it has licensed for years. FortuneDatabricks is raising at $188B to sell the build layer. Coatue is leading a round funding governance, agent tooling, and an agent-native database. BloombergGoogle shipped a cheaper, more token-efficient Gemini Flash. Output price dropped to $7.50 per million tokens, with up to 65% fewer tokens on a coding benchmark. MarkTechPost | The Overlooked The meter, not the model, is the budget risk. Uber blew a full-year AI budget in four months because coding agents price on tokens, not seats. FortuneOpen weights just made "build and self-host" competitive. Kimi K3, DeepSeek V4, and Inkling put frontier-class models at serving cost, a new column in the build spreadsheet. MarkTechPostBuild only flips if you already have the data layer. Starbucks and Uber have thousands of engineers; the median enterprise does not. TechCrunchThe casualty is per-seat pricing, not the vendors. Undifferentiated licensed software is what AI undercuts first. Forbes |
| Moving Pieces Five developments worth a CIO's attention. ProductGoogle makes the workhorse cheaper while the flagship slips again Google released Gemini 3.6 Flash on July 21, cutting the output price from $9.00 to $7.50 per million tokens, using 17% fewer output tokens on one index and up to 65% fewer on the DeepSWE coding benchmark. It teased a future Gemini 4 and, again, did not ship the flagship 3.5 Pro. The enterprise read: this is the build case in miniature. Most enterprise AI work, classification, extraction, summarization, routine code, never needs a frontier model, and the cheaper the workhorse tier gets, the more of your internal tooling is affordable to run yourself. Route by difficulty and the bulk of your calls cost a fraction of the frontier rate. The flagship is for marketing; the Flash tier is for production. Deals / InfrastructureMicrosoft bets billions that regulated Europe wants AI it can unplug Microsoft expanded its Mistral partnership on July 21 with a multibillion-dollar commitment to fund Mistral data centers in Europe, built on thousands of Nvidia Vera Rubin GPUs, and will draw on that capacity for its own cloud customers. It also added Mistral Medium 3.5 and OCR 4 to Foundry, with Medium 3.5 in Copilot Studio. The enterprise read: this is the buy side fighting back on the one axis build cannot easily match, which is control. Azure will run these models cloud-connected or fully disconnected, aimed at regulated industries and data-residency rules. If you operate under EU data law, frontier AI you can run on local, air-gapped servers just became a real procurement option, and a reason to keep buying. Governance / LegalThe $1.5B settlement that priced training data A federal judge approved Anthropic's $1.5 billion settlement with authors on July 20, covering more than 482,000 books at roughly $3,000 each, the largest known copyright recovery in history. The court had earlier found that training on copyrighted books was not itself illegal; the liability was for acquiring millions of them through pirate sites. The enterprise read: provenance is now a line item, and it cuts both ways in build-versus-buy. Build on open weights and you inherit their training-data history; buy from a vendor and you should push indemnification for training-data claims onto the provider. Either way, the courts have now attached a real, nine-figure number to getting data provenance wrong. InfrastructureAMD opens Advancing AI with a direct shot at Nvidia's rack AMD's Advancing AI 2026 runs July 22-23 at Moscone, with Lisa Su expected to detail the Instinct MI450 series and the Helios rack. The MI455X carries 432GB of HBM4 and 19.6 TB/s of bandwidth, with roughly double the compute of the MI350 generation. The enterprise read: a credible second source for training silicon is what keeps your compute costs from climbing, whether you build or buy. Every hyperscaler wants leverage against Nvidia pricing and allocation, and a real MI450 ramp gives it to them. If you buy compute through a cloud partner, a competitive AMD line is how your per-GPU-hour rate stops rising. Watch for named customer commitments in the keynote, not just teraflops. WorkforceThe layoffs cite AI; the ROI doesn't back it up Tech layoffs reached 205,832 workers across 322 events in 2026, and 54% of those events explicitly cited AI or automation. Oracle cut about 21,000 roles over twelve months, Microsoft roughly 9,000, Atlassian around 1,600. Yet a survey of 350 executives found that while 80% of large enterprises piloting AI reported workforce reductions, there was no measurable correlation between the cuts and AI ROI. The enterprise read: "AI efficiency" is doing real rhetorical work in these announcements, some earned and some cover for ordinary cost-cutting. Before you model headcount savings into any build-or-buy business case, ask for the ROI number, not the layoff number. This year the gap between the two is wide. | On the Radar Eight signals, sharpened. | Compute | Nvidia released a Vera CPU white paper with SPEC CPU 2026 results. Its first custom core, claiming up to 1.8x the performance of x86 on agentic workloads, aimed straight at AMD and Intel in the data center. Forbes | | Research | The open-weight wave compressed a year of releases into two weeks. Kimi K3 (2.8T, weights July 27), DeepSeek V4 (MIT), and Thinking Machines' Inkling (Apache 2.0) all landed, putting self-hosting back in the build conversation. MarkTechPost | | Policy | The White House frontier-model review framework is expected before August 1. It would give agencies up to 30 days to assess national-security implications before a model's public release; voluntary, not a license. AI Weekly | | Policy | AI-lab federal lobbying hit a Q2 record of $3.17M combined. Anthropic spent $1.97M and OpenAI $1.2M, up 26% and 18% quarter over quarter, as the rules that will govern them get written. Quartz | | Product | Google teased Gemini 4 while shipping the cheaper Flash tier. The pattern to watch: efficiency updates ship on time, flagship reasoning models keep slipping. 9to5Google | | Security | Oak raised $60M to govern identity and permissions for AI agents. Accel, CRV, and Greylock backed a control plane for who, and which agent, is allowed to do what. Agent access is becoming its own security category. Tech Startups | | Deployment | Bespoke Labs raised $40M to test agents before they ship. Wing VC led a round for simulation and evaluation environments, a sign that agent reliability is now a funded discipline, especially for teams building their own. Tech Startups | | Deals | AI agent startups raised more than $1.8B across 12-plus deals in July. Median valuation reached $280M, up 40% from Q1, with enterprise automation agents commanding 11-15x ARR. AI Funding |
| Quick Hits The wider field, one line each. | Oak raised $60M in seed funding for AI-native identity and agent-access governance, led by Accel, CRV, and Greylock.Tech Startups | | Bespoke Labs raised $40M Series A, led by Wing VC, for pre-deployment agent testing and evaluation.Tech Startups | | Moonshot's Kimi K3 is live at 2.8T parameters with a 1M-token context window; open weights land July 27.MarkTechPost | | DeepSeek V4 is slated for release with MIT-licensed weights from day one.AIBase | | Thinking Machines shipped Inkling under an Apache 2.0 license.Digital Applied | | MiniMax's 2.7T "M3 Pro" model surfaced in press reports.Digital Applied | | Mistral Medium 3.5 and OCR 4 were added to Microsoft Foundry, with Medium 3.5 in Copilot Studio.Microsoft | | Gemini 3.5 Flash-Lite ships at $0.30/$2.50 per million tokens and up to 350 tokens per second.Google | | Nvidia's Vera CPU packs 88 custom Olympus cores, 176 threads, and 1.2 TB/s of memory bandwidth.VideoCardz | | Oracle cut about 21,000 roles over twelve months, a 13% reduction, partly citing AI.InformationWeek | | Atlassian cut roughly 1,600 jobs, about 10% of staff, to rebalance toward AI and enterprise sales.TechCrunch | | Microsoft cut around 9,000 roles at the start of FY2026 as AI infrastructure spending climbed.InformationWeek |
| The Number 8x Frontier-everything vs. a tiered stack The gap between routing every workload to a frontier model and routing by difficulty. Frontier-everything costs about $18.40 per million tokens; a tiered setup that sends easy calls to a cheaper model runs roughly $2.31, per Digital Applied's 2026 cost analysis. Same quality on the bulk of the work, one-eighth the bill. This is the number that makes the build case real: once inference is this cheap for routine work, running your own thin layer of tooling stops looking expensive. Google's cheaper Flash tier just widened the gap, but only if your stack can route. Most cannot yet, which is the most common line of waste in an AI budget this year. | Counter-Signal RiskBuilding your own is a bill, not a discount. The tidy story is that AI just made build cheaper than buy, and this week's open-weight releases only sharpened it. The uncomfortable part is what "build" actually costs. Starbucks can attempt this because it has the engineers, the data platform, and the scale to amortize the effort across thousands of stores. Strip those away and the math inverts. For the median enterprise, replacing a licensed tool means owning the roadmap, the security patching, the on-call, and the model bills forever, trading a predictable line item for an open-ended one. Cheaper open weights lower the entry price; they do nothing for the operating cost. And the open-ended cost is not tame. Uber's budget blowout was a build-side story: internal engineers using metered coding agents, not a vendor overcharging. Governance is the other trap. Agents are entering production faster than oversight, which means home-built software carries home-built risk, with no vendor to indemnify you when an agent acts wrong, and, as this week's copyright ruling shows, no vendor to absorb a training-data claim. The honest read is not "stop buying." It is that the choice is now genuinely hard, and the companies that win will be ruthless about which slice they actually build, and disciplined enough to keep buying the rest. | From the Field The build-versus-buy debate is as old as enterprise software. What's interesting this week isn't that Starbucks flipped the answer. It's that every few days another input to the spreadsheet moves. For most of that history it was settled by a boring truth: your engineers are expensive and a vendor's are amortized across a thousand customers. Buy almost always won. The tools to build keep getting cheaper at the same moment the tools to buy get less predictable. A coding agent writes the plumbing that used to take a team a quarter. A platform like Databricks rents you the governance and data layer you would never build yourself. This week a wave of open-weight models put frontier-class capability at serving cost, and Google cut the price of its workhorse tier again. Meanwhile the license you were going to renew comes with a metered bill that surprised even Uber, and the vendors are quietly repositioning around control because raw capability alone no longer holds the account. When the inputs move like this, the spreadsheet you ran two years ago is simply wrong. But here's the part I'd hold onto. The winners won't be the companies that build everything, and they won't be the ones that buy everything. They'll be the ones that know the difference. Build the thin layer that is actually yours, the workflow and data no vendor understands, and rent the rest without ego. The mistake in both directions is the same: treating a strategic question as a religious one. It isn't. It's a spreadsheet, and this week the inputs changed again. Let's get to production,
AK | | The Agentic Enterprise Know more about AI than 95% of your peers. By 7 AM. A daily AI intelligence briefing for enterprise leaders, published by Spearhead. We build AI systems that work. Strategy. Engineering. Production. Outcomes. © 2026 Spearhead. All rights reserved. |
|
