Executive Summary

Two financial results defined the week in artificial intelligence. Anthropic reported preliminary second-quarter revenue above $11.5 billion and its first operating profit, and Databricks raised $5 billion at a $190 billion valuation on a $7 billion revenue run rate. Six new AI models also launched, competing on price and openness rather than raw capability.

The week matters because it starts to answer the question buyers keep asking: does the AI business model work? The early evidence says yes, with conditions attached. Business leaders evaluating AI vendors, contracts and budgets should read the numbers carefully before acting on them.

Estimated reading time: 11 minutes

Key Takeaways

Anthropic became the first frontier AI lab to report an operating profit, though the figure is unaudited, non-GAAP and helped by a temporary compute discount.
Databricks’ $190 billion valuation confirms that enterprise AI infrastructure is now a large, revenue-generating market rather than a promise.
Model competition has moved from capability to price, licensing and specialisation — which strengthens the buyer’s negotiating position.
Introductory model pricing is a budgeting trap: Google’s new rates are set to double on 1 January 2027.
Anthropic made four coordinated moves in one week — infrastructure, acquisition talks, agent autonomy and EU compliance — that read as pre-IPO positioning.
Roughly 95% of enterprises surveyed by Cloudera have delayed AI projects, with governance and infrastructure costs cited as the main barriers.

What Happened?

Two announcements dominated the week of 10–16 August 2026.

On 13 August, Databricks closed a $5 billion funding round at a $190 billion post-money valuation. The company reported a revenue run rate above $7 billion with year-on-year growth over 80%. The round was led by Coatue, Blackstone, MGX, T. Rowe Price and Sixth Street Growth, with around 24 venture firms participating. The valuation is 42% higher than the $134 billion figure set in February 2026. Chief executive Ali Ghodsi said investor interest reached $15 billion — three times what the company accepted. Proceeds are directed at three products: Lakebase, a database for AI agents; Genie, an AI assistant for business users; and Unity AI Gateway, a tool for governing and controlling costs across multiple AI models.

On 14 August, Bloomberg reported that Anthropic’s preliminary second-quarter revenue exceeded $11.5 billion, up from $4.73 billion in the first quarter and $787 million in the same quarter of 2025 — roughly a fourteen-fold annual increase. The company reported its first operating profit, projected at around $559 million in May investor materials. Compute cost fell from 71 cents to 56 cents per dollar of revenue, and Anthropic now has more than 1,000 enterprise accounts spending over $1 million a year.

Six new models also launched during the week.

Model	Maker	Released	Price (in/out per million tokens)	Main differentiator
Gemini 3.7 Flash	Google	13 Aug	$0.75 / $3.75 (introductory)	Tuned for coding and agent work
DeepSeek V4 Pro	DeepSeek	13 Aug	$1.32 / $3.96 (peak)	Peak and off-peak billing
Grok 4.6	xAI	12 Aug	$2 / $6	Matches GPT-5.6 Sol on benchmark scores
Muse Glimmer	Meta	10 Aug	Open weights (Apache 2.0)	30B parameters, runs on a consumer GPU
GLM-5.3	Z.ai	14 Aug	Not published	1M context window, coding and agents
Qwen3.8-Max (weights)	Alibaba	13 Aug	$2 / $6 (API)	Highest-scoring open-weights model

Pricing and benchmark figures above are reported by Capital & Compute and Sitefluence. Check them against official provider documentation before signing a contract.

Anthropic separately announced four other moves: a data centre partnership with Macquarie and GIC through Theseus Infrastructure (10 August), reported talks to acquire chip-efficiency startup Decart for around $6 billion (13 August), automatic mode as the default in Claude Code for Pro, Max and Team users (14 August), and text watermarking for EU AI Act compliance (14 August).

Why It Matters

For three years the central question about artificial intelligence has been financial rather than technical. Everyone accepted the models were improving. Nobody could show that selling access to them made money.

This week produced the first evidence on that question from a frontier lab. If Anthropic’s numbers hold, the argument that AI companies are burning cash with no route to profit becomes harder to make. That changes how a business should think about vendor risk. Signing a three-year contract with a supplier that may not exist in three years is a different decision from signing with one that covers its own costs.

The Databricks round matters for a related reason. A $190 billion private valuation supported by $7 billion of actual revenue tells you that enterprises are already paying real money for the plumbing beneath AI projects. That is spending, not sentiment.

The signal beneath both numbers: money is now moving from AI experiments into AI infrastructure. Where a company spends its budget says more about its intentions than any strategy deck.

The model launches matter for a third reason, and it is the one most likely to affect a budget this quarter. When six credible models arrive in a week, no single vendor holds the pricing power it held a year ago.

Business Impact

Enterprise

Vendor stability has improved, and so has negotiating position. Anthropic’s results reduce the risk of standardising on Claude, while the arrival of comparable models from Google, xAI, DeepSeek, Meta, Alibaba and Z.ai gives procurement teams genuine alternatives to name in a negotiation.

Databricks’ product focus is worth noting. Unity AI Gateway exists because enterprises now run several models at once and struggle to control the cost and governance of doing so. When a company raises $5 billion and spends part of it on multi-model governance, it is telling you what its largest customers are complaining about.

Small and Mid-Sized Businesses

The open-weight releases are the relevant development here. Meta’s Muse Glimmer runs on a consumer graphics card under an Apache 2.0 licence, which puts a capable model inside a small business’s own network with no per-token bill and no data leaving the building. That suits sensitive work — client records, internal documents, anything a business would rather not send to an external service.

The trade-off is real. Self-hosting means someone has to run it, update it and secure it. For most small teams a paid API remains cheaper once staff time is counted.

Software Development

Claude Code’s automatic mode is now on by default for paying users. Anthropic reports the system catches 89% of harmful actions, against 13.6% for human review, and notes that users previously approved 97% of permission prompts anyway. In plain terms: the AI coding agent no longer stops to ask before most actions.

That removes friction and adds exposure. Roughly one harmful action in ten still gets through, and it now gets through without a person seeing it first.

Compliance and Legal

Anthropic’s text watermarking marks AI-generated output automatically on models released after 2 August, with C2PA metadata for files, and the mark survives copy and paste. For businesses serving EU customers, this addresses the transparency requirements of the EU AI Act. Separately, the EU’s Digital Markets Act ruling requires Google to open Android to rival AI assistants by August 2027, with fines of up to 10% of global turnover at stake.

Key Changes

Changes with a direct effect on business decisions this quarter:

AI vendor economics are proven at one company, not the industry. Anthropic is profitable on a non-GAAP basis; OpenAI projects a $14 billion loss for 2026.
Compute costs are falling. Anthropic’s cost per dollar of revenue dropped from 71 to 56 cents, which usually flows through to customer pricing eventually.
Model pricing is diverging. Google is cutting introductory rates, DeepSeek is raising peak-hour rates, and xAI has raised cached input pricing.
Open weights are close to frontier quality. Qwen3.8-Max is now the highest-scoring open-weights model available.
Agent autonomy is becoming the default setting rather than an option a user has to switch on.
Watermarking is becoming standard for vendors selling into the European market.

Future Signal Tip

Before your next AI renewal, ask the vendor for a written price schedule covering the full contract term, not the current rate card. Introductory pricing is now common enough that a budget built on today’s published rates can be wrong by a factor of two within six months.

Industry Reaction

Coverage of Anthropic’s results was positive but heavily qualified. The Wall Street Journal compared the company’s revenue trajectory to Zoom during the pandemic and to Google and Facebook before their public listings. Reuters Breakingviews raised the sharper point: the operating profit figure excludes stock-based compensation, which at a heavily venture-funded company could be large enough to wipe out the margin on a GAAP basis. Forbes framed Anthropic and OpenAI as taking opposite routes to profitability. Analysts at SemiAnalysis estimated a second-quarter operating margin around 36% and suggested third-quarter operating profit could pass $1 billion — an outside estimate, not company guidance.

On Databricks, Bloomberg placed the round in the context of a company competing directly with Snowflake and Alphabet. The $190 billion valuation now exceeds Snowflake’s public market capitalisation of roughly $116 billion, which is a striking comparison given one figure comes from a private transaction and the other from a public market.

On the model releases, O’Reilly’s analysts made the most useful observation of the week: the more important development was not that the frontier moved, but that near-frontier capability became considerably cheaper and easier to control. An enterprise AI newsletter made a similar point differently, arguing that the strategic question has shifted away from which model to use and towards which processes a business can safely automate with measurable returns.

Commentary on Anthropic’s infrastructure deal was more sceptical. One analysis argued that frontier labs now treat dependence on a handful of cloud providers as a strategic weakness, and advised businesses to treat their model provider as a swappable component rather than a foundation.

Opportunities

The week creates several practical openings.

Opportunity	Who benefits most	What it looks like in practice
Better contract terms	Any business renewing an AI contract	Use competing models as leverage points in pricing discussions
Multi-model routing	Mid-sized and large teams	Send simple tasks to cheap models, complex reasoning to premium ones
Self-hosting for sensitive data	Regulated businesses, professional services	Run open-weight models internally so client data never leaves the network
Faster agent-based development	Software teams	Less time approving prompts, more work completed per developer hour
EU market confidence	Businesses selling into Europe	Watermarked output reduces the compliance burden of AI-assisted content

The routing opportunity is the one most businesses underuse. A large share of everyday AI work — summarising, classifying, drafting standard replies — does not need a premium model. Splitting that traffic can cut a monthly bill substantially without any noticeable drop in quality.

Risks & Limitations

The headline numbers deserve scrutiny.

Anthropic’s profit carries four caveats. The figures are preliminary and unaudited. They are non-GAAP, excluding stock-based compensation while including training costs. A compute discount from SpaceX artificially lowered second-quarter costs. And Anthropic itself has warned that profitability in the third and fourth quarters is not guaranteed.

Databricks’ valuation is a private-market price. It reflects what a small group of investors agreed to pay in a negotiated transaction, not what a public market would bear. The company has not disclosed a full capitalisation table or the terms attached to the securities.

Model benchmarks should not be taken at face value. Meta’s Muse Glimmer showed the widest gap between vendor-reported and independently measured scores of any release this month. Vendor benchmarks for GLM-5.3 and Qwen3.8-Max have not been independently reproduced.

Licensing needs checking before commercial use. Qwen3.8-Max uses a custom licence with revenue gating above $50 million. Open weights do not automatically mean free commercial use.

Agent autonomy shifts risk. An 89% catch rate for harmful actions means roughly one in ten is missed, now without human review in the loop.

The Decart acquisition may not happen. Talks were reported but not finalised, and both companies declined to comment.

Accuracy note: Anthropic’s second-quarter figures come from reported investor materials rather than audited accounts. OpenAI’s S-1 had not been publicly filed as of 16 August 2026, so all OpenAI financial figures referenced here are reported rather than filed. Deal value and capacity for the Theseus Infrastructure partnership were not disclosed. Model pricing was sourced from secondary trackers, not directly from every provider’s documentation.

Future Outlook

Confirmed. Google’s Gemini 3.7 Flash introductory pricing doubles on 1 January 2027. The EU requires Google to open Android to rival AI assistants by August 2027. Anthropic’s watermarking applies automatically to models released after 2 August 2026.

Likely over the next 6–24 months. Databricks is widely expected to pursue a public listing in late 2026 or 2027. Anthropic is reportedly targeting an October 2026 listing, though this is not officially confirmed. Competitors are likely to follow Anthropic’s lead on default agent autonomy. Watermarking will probably become standard across vendors serving European customers.

Editorial interpretation. Expect model pricing to become harder to compare, not easier. Peak and off-peak billing, context-length thresholds and capability tiers all make headline rates less meaningful. Businesses will need to model their actual usage patterns rather than compare list prices. We also expect the capability gap between open and closed models to keep narrowing, which will put steady downward pressure on mid-tier pricing.

The Future Signal

The industry has spent three years arguing about whether AI companies can make money. That argument is quietly ending, and a more useful one is starting: which parts of the AI supply chain actually capture value?

Look at where the money went this week. Databricks raised billions for agent databases and multi-model governance. Anthropic partnered on data centres and opened talks to buy a chip-efficiency company. Neither move is about making models cleverer. Both are about controlling costs and controlling the layer beneath the model.

That tells business leaders something specific. The model itself is becoming a commodity — interchangeable, competitively priced, available in open-weight form at close to frontier quality. The durable value is accumulating one layer down, in the infrastructure that runs models cheaply, and one layer up, in the systems that govern how a business uses them.

For most companies the practical consequence is simple. Do not build your AI strategy around a model. Build it around the work you want done, and treat the model as a part you can replace when a better-priced one arrives. That has always been sound advice. This week it became measurably cheaper to follow.

What Businesses Should Do Next

Recommendations in priority order:

Review your AI contract renewal dates. With six credible models released in one week, competition is at its strongest point yet. Renewals in the next two quarters should be negotiated, not rolled over.
Model your costs on post-introductory pricing. Assume introductory rates end. For anything running on Gemini 3.7 Flash, budget on the January 2027 rates.
Test one cheap model on one high-volume task. Pick a repetitive job — summarising support tickets, classifying enquiries, drafting standard responses — and run it on a lower-cost model for a month. Compare quality and cost honestly.
Review your agent permissions. If your development team uses Claude Code, automatic mode is now on by default. Confirm that is what you want, and check what an agent can reach in your systems.
Check licences before self-hosting. Open weights are not the same as unrestricted commercial rights. Read the terms, particularly revenue thresholds.
Monitor rather than act on the IPO stories. Anthropic’s and Databricks’ listings will generate a great deal of noise. Neither changes what a model does for your business today.

Quick Review Checklist

Renewal dates for all AI contracts documented
Budget rebuilt on post-introductory pricing
One high-volume task identified for a cheaper model trial
Agent permissions and default settings reviewed with the development team
Licences checked for any open-weight model in commercial use
EU AI Act transparency requirements confirmed with whoever owns compliance

Brief Signals

IPO and financial

OpenAI’s public prospectus was reported to be days away from filing, with roughly $2 billion in monthly revenue, a projected $14 billion loss for 2026 and a reported $852 billion valuation. None of these figures had been filed as of 16 August.
Lovable raised $400 million at a $13.3 billion valuation, more than doubling its worth in eight months on a reported $500 million in annual recurring revenue.

Enterprise partnerships

IBM and OpenAI announced a partnership combining OpenAI models with IBM’s consulting and governance services for enterprise deployment.
OpenAI expanded its Daybreak cybersecurity service into Blue and Red tiers. Its GPT-5.6-Cyber model scored 95% on cybersecurity prompts against 1.5% for the default model, and reportedly identified two previously unknown bugs in the V8 engine.

Regulatory

The EU’s Digital Markets Act ruling requires Google to open Android to rival AI assistants by August 2027, with penalties of up to 10% of global turnover.

Market milestones

Google’s Gemini app reached one billion monthly active users, making it the company’s fourteenth product to do so. ChatGPT passed the same mark in June.
Koray Kavukcuoglu replaced Demis Hassabis as SVP at Google DeepMind, a change widely read as prioritising execution over long-horizon research.
A Cloudera survey found 95% of enterprises had delayed AI projects, citing data governance and infrastructure costs as the main barriers.

That last figure deserves attention alongside the week’s headline numbers. Vendor revenue is growing quickly while most enterprises report that their own projects are stalled. Someone is spending heavily. It is not the average company.

Frequently Asked Questions

Does Anthropic’s operating profit mean AI companies are now profitable? It means one company reported one profitable quarter on a non-GAAP basis. The figure excludes stock-based compensation, the numbers are unaudited, and a temporary compute discount lowered costs. It is encouraging evidence, not settled proof.

Should we switch AI vendors based on this news? Not on its own. Vendor financial health is one factor among several, alongside model quality, data handling, integration effort and switching cost. The stronger move is using the competitive market to renegotiate rather than to migrate.

What is an open-weight model, and why does it matter to my business? It is a model whose underlying files are published, so you can run it on your own hardware instead of calling an external service. It matters when data cannot leave your network, or when API costs at high volume exceed the cost of running it yourself.

Is running a model ourselves cheaper? Only above a certain volume. Below it, staff time for setup, security and maintenance usually costs more than the API bill. Estimate both before deciding.

What does “introductory pricing” actually cost us? It costs you budget accuracy. Google’s Gemini 3.7 Flash rates double on 1 January 2027. If your annual budget assumes current rates, it is wrong for the second half of that year.

Should we be concerned about AI coding agents running without approval prompts? Concerned is too strong; deliberate is right. Auto mode reflects the fact that users approved almost every prompt anyway. The sensible response is to review what your agents can access rather than to switch the feature off.

How does watermarking affect content our business produces with AI? Output from newer Anthropic models carries a mark that survives copying and pasting. It supports EU AI Act transparency requirements. If you publish AI-assisted content, discuss disclosure with whoever handles compliance.

Which single change from this week should we act on first? Contract renewals. Competition is at its strongest, and pricing is the one area where this week’s news translates directly into money saved.

Conclusion

The week’s most important development was not a new capability. It was evidence that the money behind artificial intelligence is beginning to work — with enough caveats attached that no one should treat it as settled. Anthropic showed a frontier lab can turn an operating profit. Databricks showed enterprises are paying seriously for the infrastructure underneath.

The practical takeaway is narrower than the headlines. Competition among model providers is now genuine, which makes this a good quarter to review contracts, test cheaper options on routine work, and check what your AI agents are permitted to do.

Over the next year, expect model access to keep getting cheaper and the interesting decisions to move elsewhere: to governance, to cost control, and to choosing which work is worth automating at all. Build for that, and a change of model becomes a purchasing decision rather than a rebuild.

Anthropic Reports First Operating Profit as AI Spending Accelerates