Today's 12 Stories — Friday, August 14, 2026

Funding & Investment Story 1 of 12

Databricks Raises $5 Billion at a $190 Billion Valuation and Names Its Agent Bottleneck

Databricks said on Wednesday that it has raised $5 billion at a $190 billion valuation, in a round led by Coatue with participation from Blackstone, MGX, T. Rowe Price, Sixth Street Growth, BOND, Clearlake Capital, Point72, Premji Invest and TPG alongside existing backers. The company disclosed alongside the raise that its overall revenue run rate has surpassed $7 billion, growing more than 80 percent year over year, and that it has been free cash flow positive on an adjusted basis over the last twelve months.

The growth figures matter more than the valuation, because they arrive with a segment breakdown that tells executives where enterprise money is actually moving. Databricks said its lakehouse revenue run rate now exceeds $1.5 billion and is growing over 100 percent year over year. Lakebase, the serverless Postgres product it launched barely a year ago, is already exceeding a $100 million revenue run rate. More than 1,000 customers now consume at over $1 million a year, and more than 100 consume at over $10 million.

That last pair of numbers is the one to sit with. A base of 100 customers each spending eight figures annually on data and AI infrastructure is not a pilot economy. It is a set of companies that have already crossed from experimentation into production dependence, which changes both their negotiating position and their switching cost.

The company framed the round around a specific technical claim. Chief executive Ali Ghodsi said enterprises want agents with context, accuracy and budget control, and named Lakebase, the Genie assistant and the Unity AI Gateway as the products meant to deliver those three things together. The framing is a quiet admission of what has gone wrong in enterprise agent deployments so far. Agents fail in production less because the underlying model is weak and more because the agent cannot reach governed data quickly enough, cannot be audited when it is wrong, and cannot be capped when it starts spending.

Coatue co founder Thomas Laffont, whose firm led the round, described Databricks as compressing research and development timelines from years to months and operating like a research lab. Investors have been saying versions of that about data platform companies for a decade. What is new is that the revenue mix now supports it: the fastest growing lines are the ones sold specifically to teams putting agents into production, not the ones sold to analytics teams building dashboards.

For buyers, the strategic read is about consolidation pressure. A vendor with a $7 billion run rate, positive cash flow and $5 billion of fresh capital is not going to stay in its lane. Databricks is now simultaneously a data warehouse, an operational database, a governance layer and a model gateway. Every one of those was a separate procurement conversation two years ago. Companies still buying them separately should expect their incumbent vendors to start defending those seats hard, and should ask what happens to their leverage when four contracts become one.

AI Models Story 2 of 12

Grok 4.6 Arrives at Two Dollars In, and Its Reasoning Dial Defaults to High

xAI has released Grok 4.6, positioning it in its own developer documentation as the flagship model for code and everything else, with agentic tool calling, minimal hallucinations and configurable reasoning. The published pricing is $2.00 per million input tokens and $6.00 per million output tokens. The context window is 500,000 tokens, and the model carries a knowledge cutoff of February 1, 2026.

The headline number is the price. At two dollars in and six dollars out, Grok 4.6 lands materially below where frontier tier pricing sat a year ago, and it does so with a context window large enough for most document heavy enterprise workloads without resorting to retrieval tricks. For teams whose agent costs are dominated by long input contexts rather than long generations, the three to one ratio between input and output pricing is worth modelling directly against current spend.

The less discussed detail is in the reasoning documentation, and it is the one likely to show up on invoices. xAI documents a reasoning effort parameter with four values: low, medium, high and xhigh. If a request does not specify one, the default is high. That is a deliberate design choice in favour of answer quality, and it is defensible. It also means that any team that has migrated an existing integration to Grok 4.6 without touching its request payload is now paying for high effort reasoning on every call, including the classification, extraction and routing calls that never needed it.

The xhigh tier is new to this generation. xAI documents it as available on Grok 4.6 and later; on models that do not support it, such as Grok 4.5, a request for xhigh is served at high rather than rejected. That silent downgrade is convenient for portability and dangerous for benchmarking, because a team comparing 4.5 and 4.6 at xhigh is not actually comparing the same setting. Anyone running a model bake off should confirm which effort level each side actually executed rather than trusting the parameter they sent.

xAI also documents that presence penalty, frequency penalty and stop sequences cannot be used with its reasoning models. Requests that include them return an error rather than silently ignoring them. Teams porting prompt configurations from older non reasoning deployments should expect those calls to fail outright, which is the better failure mode but still a migration task somebody has to own.

Taken together, the release fits the pattern that has defined this year in frontier models. The capability gap between the top handful of models keeps narrowing while the price and configuration surface keeps widening. The competitive question for a buyer is no longer which model is smartest. It is which vendor's defaults, error behaviour and effort tiers match the shape of the work, because that is where the difference between a reasonable bill and an unpleasant one now lives.

Industry Dynamics Story 3 of 12

Lenovo's AI Revenue Reaches $9.3 Billion and Its Server Business Nearly Doubles

Lenovo reported first quarter fiscal 2026/27 group revenue of $26.9 billion, up 43 percent year over year, in what the company called its strongest quarter in group history. AI related revenue reached $9.3 billion, or 35 percent of total revenue, growing 60 percent year over year. Adjusted non HKFRS net income came in at $1,075 million, up 176 percent.

The segment detail is where the story changes shape. The Infrastructure Solutions Group, which sells servers and data centre systems, posted $8.5 billion in revenue, up 98 percent year over year, at a 9.1 percent operating margin. The Intelligent Devices Group, the PC and smartphone business that has been Lenovo's identity for two decades, grew 27 percent to $17.1 billion at a 7.1 percent margin. The Solutions and Services Group reached $2.9 billion, up 28 percent, at a 24.2 percent margin.

A server business growing at 98 percent while earning a higher operating margin than the flagship device business is not a rounding adjustment. It is a change in what kind of company Lenovo is. For most of its modern history the infrastructure segment was the strategically important, structurally unprofitable part of the portfolio, subsidised by devices. That relationship has now inverted on growth, and the margin gap has closed to two points.

Executives should read one number here carefully. On a statutory basis, profit attributable to equity holders was a loss of $609 million, against a $505 million profit a year earlier, even as adjusted income tripled. The gap between the two figures is large enough that anyone using Lenovo's results as a proxy for the health of the AI server market should be explicit about which measure they are quoting. Adjusted income is the one the company leads with, and the one that supports the growth narrative. The statutory line is the one that shows what carrying this build out currently costs.

Research and development spending rose 30 percent, which is the ordinary consequence of trying to hold share in a segment where the compute vendor sets the pace and the reference designs change every few quarters. Server assemblers competing for hyperscaler and neocloud business are in a genuinely difficult position: demand is enormous, the input costs are set by a supplier with pricing power, and differentiation has to come from integration, power engineering and delivery speed rather than silicon.

The broader signal for buyers is about supply. When a vendor of Lenovo's scale nearly doubles infrastructure revenue in a year, it is absorbing allocation that used to reach the mid market. Enterprises planning on premise AI capacity for the next two quarters should be checking lead times now rather than assuming that last year's procurement timelines still hold. The companies with the most negotiating leverage in this market are the ones who committed early, and that window is closing rather than opening.

AI Research Story 4 of 12

Google's Medical AI Runs Live Video Consultations and Scores Level With Doctors

Google Research has published results from a study of AMIE, its research medical AI system, conducting real time video consultations. The design was a multi arm randomised objective structured clinical examination covering 100 clinical scenarios across five body systems, run as 300 live consultations with 15 trained patient actors and assessed by an independent panel of 20 experienced primary care physicians.

Three arms were compared: AMIE running over live video, AMIE in a text only configuration, and a control arm of 10 board certified primary care physicians using the same video interface. Clinical evaluators rated the video version of AMIE on par with the physicians across history taking thoroughness, diagnostic accuracy, management appropriateness and communication quality. On the specific task of eliciting physical signs and guiding a virtual examination, the video system scored significantly higher on average than both the physicians and the text version of AMIE.

That last finding is the interesting one, and it is not the finding most coverage will lead with. Guiding a remote patient through a self examination is a skill that video telehealth has always struggled with, because it requires sustained, patient attention to what the camera is showing and an unhurried willingness to ask someone to move the phone and try again. It is exactly the kind of work where a system with no time pressure and no queue behind it has a structural advantage over a clinician with a full afternoon list.

The architecture is worth noting for anyone building agents on live sensory input. Google Research describes AMIE as built on Gemini and Project Astra, using an asynchronous multi agent design that splits the work between a Talker agent handling conversation, a Planner agent handling clinical reasoning, and a Perception agent handling the visual stream. Decoupling the conversational loop from the slower reasoning and perception loops is a general pattern, not a medical one, and it is a reasonable template for any real time multimodal agent where latency and depth pull in opposite directions.

The limitations Google states are substantial and stated plainly. The study ran entirely with professional patient actors in simulated settings, not real patients. Scenarios were confined to conditions that can be authentically portrayed through acting, which excludes a great deal of what actually walks into a clinic. The system showed occasional perceptual and reasoning errors and intermittent technical problems.

For health system executives the practical read is narrow but real. This is not evidence that an AI can replace a video consultation with a doctor. It is evidence that the physical examination step of remote care, long treated as the reason telehealth cannot go further, may be more tractable than assumed. Organisations running virtual first care should be watching whether that specific capability holds up outside a simulation, because it is the constraint that has capped what remote consultation could safely cover.

AI Infrastructure Story 5 of 12

Vantage Data Centers Weighs a Hundred Billion Dollar Listing

Reuters reported on Thursday that Vantage Data Centers, backed by private equity firm Silver Lake and infrastructure investor DigitalBridge Group, is exploring an initial public offering at a valuation of about $100 billion, or alternatively a sale, potentially as soon as next year. According to that reporting, an IPO could raise around $10 billion. Vantage has raised roughly $11 billion since late 2023, including a $9.2 billion equity investment led by DigitalBridge and Silver Lake. The company has met informally with financial advisers but has not launched a formal process, and the deliberations remain at an early stage.

The number to hold onto is the ratio. A company that raised roughly $11 billion in under three years is being discussed at a valuation approaching ten times that. Data centre capacity has become the scarcest input in the AI build out, and the market is now pricing developable, powered land and grid interconnection the way it once priced spectrum licences.

Vantage recently partnered with Oracle and OpenAI on a data centre campus in Wisconsin connected to Stargate, the joint venture targeting $500 billion of infrastructure investment and 10 gigawatts of capacity. That relationship is the clearest explanation of why a listing is being discussed now rather than in three years. A developer with signed anchor commitments from the largest compute buyers in the market has a contracted revenue profile that public investors can underwrite, which is a very different asset from a speculative development pipeline.

For enterprise leaders the relevance is not the deal itself but what it reveals about the capital structure forming underneath their compute costs. When data centre developers list publicly at these multiples, the cost of capacity stops being set by construction economics and starts being set by capital market expectations. Public shareholders will want returns commensurate with a hundred billion dollar valuation, and those returns come from lease rates paid by the companies renting the capacity, which flow through to the price of inference.

There is also a concentration question worth asking internally. The same handful of hyperscalers and AI labs are anchoring most of the large campus projects being financed right now. That makes the developers investable and simultaneously makes them dependent. If any large anchor tenant materially slows its capacity commitments, the effect will not be confined to that tenant's balance sheet.

Nothing here is decided. Reuters described the discussions as early and subject to change, and a sale remains on the table alongside a listing. But the direction is consistent with everything else visible in this market this month. Capital keeps finding its way to the physical layer, at valuations that assume the demand curve does not bend.

Industry Dynamics Story 6 of 12

Anthropic Lines Up Its Largest Acquisition, and the Target Is an Efficiency Play

Bloomberg reported on Thursday that Anthropic is in talks to acquire Decart, an Israeli artificial intelligence startup, for about $6 billion. If completed it would be the largest acquisition Anthropic has made. According to that reporting, Decart was valued at about $4 billion following a funding round in May of this year, up from $3.1 billion in August 2025, with investors including Radical Ventures as lead, alongside Nvidia, Atreides Management, Valor Equity Partners, Adobe Ventures, Sequoia Capital, Benchmark and Zeev Ventures.

Decart builds world models that simulate physical environments, together with chip efficiency software. Its consumer facing work includes processing live video feeds to generate real time modified video, with applications such as virtual garment try on for fashion commerce. That product surface is not obviously adjacent to Anthropic's business, which is why the reported rationale is the interesting part.

The reported logic is that Decart's video simulation and chip efficiency technology would go into Anthropic's inference team, helping existing infrastructure absorb more demand while reducing training costs through more efficient chip utilisation. Read plainly, this is not a product acquisition. It is a margin acquisition.

That framing deserves attention from anyone buying frontier model capacity. The competitive frontier among labs has quietly shifted from raw capability to unit economics. Every lab can point to benchmark parity within a narrow band. What separates them commercially is how many tokens they can serve per dollar of compute, because that determines whether they can cut prices without cutting into gross margin, and whether they can serve enterprise contracts at volume without rationing capacity. A lab that is willing to spend six billion dollars for inference efficiency is telling the market where it believes the binding constraint is.

There is a second signal in the price. Six billion dollars for a company most recently valued around four billion is a substantial premium for a business whose visible products are in adjacent consumer categories. Premiums of that size are usually paid for scarcity of talent and technique rather than for revenue. The market for people who can materially improve inference throughput on existing silicon is very small, and every lab is bidding for the same names.

Executives negotiating multi year model contracts should factor this in. Inference costs across the frontier have fallen consistently, and vendors have passed a meaningful share of that through to list prices. Deals like this are the mechanism behind those cuts. It is reasonable to negotiate for price step downs over a contract term rather than accepting today's rate card as fixed, because the vendors are actively engineering their costs downward and expect to keep doing so.

The talks are reported, not confirmed, and may not result in a transaction.

AI Infrastructure Story 7 of 12

L&T Wins India's Largest Nvidia B300 Build for an American AI Cloud

Larsen & Toubro announced on Thursday that it has secured a mega order, as part of a strategic partnership with Together AI, to build what it describes as India's largest Nvidia B300 AI factory. The facility will house 10,000 Nvidia B300 GPUs at the Chennai data centre campus operated by Vyoma, an L&T company, with LTN Compute involved as the group's AI infrastructure arm. The first phase is designed for 250 megawatts, with 150 MVA of power infrastructure readiness. Under L&T's own disclosure convention, a mega order falls between 10,000 crore and 15,000 crore rupees.

The customer detail is the part worth pausing on. Together AI is a United States based AI cloud platform, and it is contracting for training, fine tuning and large scale inference capacity in Chennai rather than in North America or Europe. That is a routing decision driven by the two inputs that now govern where AI capacity gets built: available power and time to energisation. India has been aggressively courting exactly this workload, and a 250 megawatt first phase with power infrastructure already provisioned is a credible answer to the constraint that has stalled projects elsewhere.

L&T chairman and managing director S N Subrahmanyan framed the deployment as a milestone in the country's AI build out. Together AI co founder and chief executive Vipul Ved Prakash described making AI globally accessible as the largest infrastructure build out in human history, and credited L&T with understanding that. The rhetoric is standard. The engineering commitment behind it is not: a conventional engineering and construction group is taking on the electrical, thermal and delivery risk of a gigawatt class programme, starting with a quarter gigawatt.

For executives the strategic implication is about where inference will physically live over the next three years. Capacity is migrating toward jurisdictions that can deliver power quickly, and the resulting map does not match the map of where enterprise data currently sits. That matters for latency on interactive agent workloads, and it matters considerably more for data residency. A European or American company buying inference from a provider whose newest capacity is in South Asia should be asking which region its requests will actually be served from, and whether it has contractual control over that.

It also marks a shift in who builds this infrastructure. The winners of large AI data centre contracts are increasingly heavy engineering firms rather than specialist data centre developers, because at this scale the hard problems are substation construction, cooling at density and schedule certainty rather than rack design. Enterprises negotiating colocation for their own AI capacity should weight delivery track record accordingly, since the constraint that will delay a project is almost never the compute.

Policy & Regulation Story 8 of 12

The AI Watchdog Debate Moves Indoors, and Treasury Has Its Own Draft

The proposal Demis Hassabis made publicly in July for an industry funded, government overseen AI standards body has, according to reporting published this week by the Wall Street Journal and Bloomberg, been carried into private meetings with senior United States officials, including Treasury Secretary Scott Bessent and White House Office of Science and Technology Policy director Michael Kratsios. Bloomberg reported that Bessent independently developed a comparable proposal on the same regulatory model, now under review by White House Chief of Staff Susie Wiles.

The public version of the Hassabis proposal is on the record and specific. Speaking to Axios in July, he described a body modelled on FINRA, the self regulatory organisation that oversees United States broker dealers, with a majority independent board. Frontier labs would share models voluntarily up to 30 days before release for testing of dangerous cyber, biological and deception capabilities, and the body would be able to coordinate industry wide slowdowns if risks escalated. He argued that once the mechanism proved itself, formalisation requiring models to pass before United States deployment could quickly follow, and he wanted the structure standing before the end of the year.

The convergence is what is new. When the most prominent technical advocate of a regulatory architecture and a cabinet secretary arrive at the same institutional design independently, the probability of something being built rises sharply, and the window for affected parties to shape it narrows.

Executives should understand what the FINRA analogy actually implies, because it is not a soft option. Self regulatory organisations of that type are funded by the industry, staffed largely from it, and possess real enforcement authority delegated by statute. They tend to produce rules that are technically sophisticated, expensive to comply with, and structurally favourable to incumbents who can absorb the compliance overhead. A 30 day pre release review period is trivial for a lab that ships four models a year and materially disruptive for a startup that ships continuously.

The second order effect lands on enterprise buyers rather than model developers. A pre deployment testing regime introduces a gate between a model being finished and a model being available. Organisations that have built roadmaps around rapid access to each new frontier release should think about what a mandatory month of latency does to those plans, and about whether their vendor contracts contain any commitment regarding availability timelines that a regulator could delay.

Nothing has been proposed formally, no legislation exists, and both the Hassabis effort and the Treasury work are at the stage where design choices are still open. That is precisely why this is the moment for organisations with a serious stake in model availability to be in the conversation. Once an institution of this kind exists, its rules are set by whoever showed up while it was being drawn.

Industry Dynamics Story 9 of 12

Microsoft Keeps Shrinking in China While Its AI Business There Grows

Reuters reported on Thursday that Microsoft has closed at least 15 branch offices and joint ventures in China over the past five years, while China now accounts for about 1.5 percent of the company's global revenue as of 2024. The reporting also describes a 2024 offer to relocate 1,000 of its top engineers out of the country, which about a third accepted, and a shift of Microsoft Research Asia work toward labs in Vancouver, Singapore and Tokyo.

What Microsoft is keeping is more informative than what it is closing. According to the reporting, the company continues to provide Azure services to Chinese companies serving Western markets, and continues to support ByteDance's overseas operations. Quantum computing and other sensitive research in China has been wound down, and direct participation in government procurement has effectively ended.

The resulting shape is a deliberate one, and it is worth naming precisely because it will become a common template. Microsoft has not left China and shows no intention of doing so. It has instead reduced its exposure to the domestic Chinese market, which is where the regulatory and political risk concentrates, while retaining the part of the business that serves Chinese companies operating internationally, which is where the revenue and the AI growth are. Former China head Alain Crozier told Reuters that geopolitics made some days harder but that there was never a crisis, and that the company never changed its commitment to bringing technology into China.

For multinationals with meaningful China operations, this is a case study in partial decoupling that neither exits nor pretends conditions have not changed. The organising question is not whether to be in China. It is which specific activities carry political risk that no commercial return justifies, and which do not. Government procurement and sensitive research fall on one side of that line. Cloud services for export oriented Chinese customers fall on the other.

The talent figures deserve separate attention. Offering relocation to 1,000 senior engineers and having roughly two thirds decline is not a failed programme, but it is a clear result. Deep technical talent has ties that survive corporate restructuring, and any firm modelling a similar move should assume a substantial share of the people it most wants to keep will choose to stay. Microsoft's reported attrition improvement, from around 17 percent in the mid 2010s to under 10 percent, came from building new business lines locally rather than from relocation.

The broader point for boards is that the China operating model most companies designed before the export control era no longer exists, and the ones that adjusted early did so by segmenting activities rather than by making a single binary decision. Nearly a decade of restructuring produced a footprint of 1.5 percent of revenue that Microsoft still considers strategically worth holding.

AI Infrastructure Story 10 of 12

Kioxia and Sandisk Push Flash to 4.8 Gigabits a Second Because Inference Is a Storage Problem

Kioxia and Sandisk announced on Wednesday a ninth generation 2Tb QLC 3D flash memory technology aimed at AI and data intensive workloads. The headline specification is a NAND interface speed of 4.8 gigabits per second, which the companies describe as a 33 percent improvement over their eighth generation devices. The design uses a six plane architecture for higher write and read bandwidth, better power efficiency in both directions, and CMOS directly Bonded to Array construction, which pairs an advanced CMOS wafer with the existing memory array platform.

The specifications are incremental. The reason they matter is not. For most of the last three years the public conversation about AI infrastructure constraints has been about GPUs, and more recently about power and high bandwidth memory. Storage has been treated as solved. It is not, and the reason is the shape of modern inference workloads.

Retrieval augmented generation, long context serving and agent systems that maintain persistent state all convert what used to be a compute bound problem into a data movement problem. A model with a very large context window is only useful if something can feed it. Key value cache offload, vector index serving and checkpoint loading are all storage bandwidth problems, and QLC flash is where the cost per terabyte makes them economically viable at scale. A third more interface bandwidth on high density QLC changes how many tokens per second a given rack can actually serve, independent of the accelerator in it.

The commercial context is equally relevant. Memory and storage suppliers have spent this cycle allocating capacity toward AI customers who will pay for it, and enterprises buying conventional storage have already felt that in pricing and lead times. A new generation aimed explicitly at AI infrastructure signals that the allocation preference is hardening rather than easing.

Kioxia did not publish quantitative figures for bit density, program throughput or read latency in this announcement, describing several improvements qualitatively as higher or better, and gave no sampling or production dates. That is normal for a technology announcement at this stage and it is also a reason not to build a procurement plan around it yet. The interface speed figure is the one concrete number available.

For infrastructure leaders the practical action is to check whether storage is actually characterised in their AI capacity planning. Many organisations sized their AI infrastructure on GPU count and memory, and treated storage as whatever the reference architecture specified. Teams running retrieval heavy or agent workloads in production frequently discover that the bottleneck sits in the data path rather than the accelerator, at which point they have bought the expensive component and underprovisioned the cheap one.

Enterprise AI Story 11 of 12

Okta Attacks Agent Cost at the Identity Layer Instead of the Model Layer

Okta announced on Thursday a mechanism for cutting AI agent costs that does not touch the model at all. The approach applies identity scoping to Model Context Protocol tool lists, filtering which tools are visible to an agent based on the permissions of the user the agent is acting for, before the model ever processes the request. The company said internal modelling showed tool visibility reduced by more than 90 percent in some scenarios, with tool schema token costs falling by roughly the same margin.

The mechanic is simple enough to explain in a sentence, and most teams running agents in production have not done it. Every MCP tool an agent can call has a schema, and every schema occupies tokens in the context of every single request, whether or not the agent uses it. An enterprise agent connected to a few dozen internal systems can be spending a meaningful share of its input budget describing capabilities that the person on whose behalf it is acting has no authority to invoke.

Filtering that list by identity fixes two problems with one change. The cost problem is obvious. The security problem is less discussed and more serious: a tool that is described to the model is a tool the model can attempt to call, and prompt injection attacks work considerably better when the injected instruction can reference a real, described capability. Removing unauthorised tools from the context removes them from the attack surface rather than relying on a downstream authorisation check to catch the call.

Paul Webber, principal cybersecurity industry analyst at Software Analyst Cyber Research, said cost control for agents is best provided using identity governance tools that offer more granular control and precision without disrupting business processes. The framing is worth taking seriously by procurement teams, because it relocates a problem currently being handled badly. Most organisations are trying to control agent spend through model selection, prompt compression and rate limits, all of which trade quality for cost. Scoping tools by identity reduces cost without reducing capability for any user, because no user loses access to anything they were entitled to use.

The stated figures deserve a clear caveat, and Okta supplies it. The analysis rests on internal modelling using product data and public documentation, not on measured customer deployments, and no absolute token counts or dollar figures were provided. A greater than 90 percent reduction in tool visibility describes a scenario in which an agent had access to a very large tool catalogue and a given user was entitled to very little of it. That is a real configuration in large enterprises and it is not the median one.

The transferable lesson does not depend on the vendor. Teams running MCP connected agents should measure what proportion of their input tokens is tool schema overhead. It is a five minute measurement, most teams have never taken it, and the answer is frequently uncomfortable.

AI Business Models Story 12 of 12

Legal AI Valuations Are Doubling in Months, and Legora Is the Test Case

Swedish legal AI company Legora is seeking new funding at a valuation of more than $10 billion, according to reporting by the Financial Times published this week. The company has not published that figure itself. What Legora does state on its own newsroom is a $5.55 billion valuation from its Series D, a $550 million round that closed in March, and more than $100 million in annual recurring revenue as of its April announcement acquiring legal research company Qura.

Set those two numbers beside each other and the pattern in professional services AI becomes visible. A company that raised at $5.55 billion in March is reportedly in the market at roughly double that a few months later, on a business that had crossed $100 million of recurring revenue. Whatever multiple that implies, it is not being set by current revenue.

It is being set by a bet about substitution. Legal services are among the most attractive targets for AI displacement on paper: the work is document heavy, the output is text, the quality bar is high but assessable, and the billing model has historically charged for hours in a way that made buyers acutely conscious of every one. If a meaningful share of contract review, due diligence and legal research moves to software, the addressable revenue is enormous, because it is currently being captured at law firm rates.

Executives buying these tools should hold two things in mind. The first is that valuations at this level price in near total success, which means the vendors carrying them are under real pressure to grow into the number. That shapes pricing behaviour over a contract term, and buyers signing multi year agreements should look carefully at renewal escalators and at what happens to per seat pricing as usage expands.

The second is consolidation. Legora acquired Qura in April to build what it described as an AI native legal research platform. Point solutions in this category are being bought, and the tool a firm selects this year may be a component of a different vendor's suite next year. That argues for contract terms that survive an acquisition and for data portability provisions that are actually exercisable, not aspirational.

The underlying question none of the funding rounds answer is whether the productivity gains land with the firms or with their clients. If AI cuts the hours required for document review by a large factor, the value shows up either as higher firm margins or as lower client bills, and which one depends entirely on how competitive the market for that work is. General counsel renegotiating outside counsel arrangements over the next year have an unusually strong position, and it comes from knowing what the work now costs their law firm to produce.