DEV Community: Fathom News

MoneyGram Calls In AI To Catch Fake IDs As Deepfakes Get Cheap

Fathom News — Wed, 28 Oct 2026 11:30:31 +0000

MoneyGram, which moves money for more than 60 million people, has picked AI startup AiPrise to check customers and data across its network, as fake IDs become cheap and easy to make.

Key Takeaways

MoneyGram will use AiPrise to run some of its verification checks through one platform
AiPrise draws on more than 100 data sources worldwide and covers 200+ countries
MoneyGram serves 60 million active customers through nearly half a million retail locations
Both CEOs told Fathom the real fight is fake IDs and messy rules from country to country

MoneyGram Picks AiPrise For Fraud Checks

The two companies announced the deal on Wednesday. MoneyGram will use AiPrise for compliance and fraud prevention, starting with select verification processes in some of its markets.
The aim is to give MoneyGram's teams one consistent way to review information and record how each verification decision was made. MoneyGram operates in more than 200 countries and territories and is rebuilding much of the technology behind its network.

Fake IDs Now Take Seconds To Make

Chaitanya Sarda, CEO of AiPrise, told fathom.news that checking a document photo is no longer enough on its own.

“Generative AI can fake a passport or selfie in seconds, so any check that asks ‘does this image look real?’ is losing.”

He said the fix is to check people against sources fraudsters cannot fake, such as government registries, cross-checked data and warning signs like the same face showing up on unrelated accounts. Over time, he expects verified identities to become reusable, so customers stop uploading documents again and again.

Who Pays When An AI Lets A Crook Through

Sarda said regulators care whether a decision can be explained, not whether a person or a machine made it. He argued AI agents can clear that bar more easily than manual review because they log every source checked.

“Liability doesn’t change: the regulated institution owns its decisions, and vendors are accountable for building accurate, auditable systems.”

He expects supervisors to focus on how these AI agents are tested, monitored and overridden by humans.

Tighter Checks Without Locking Out Migrants

Many people who send money home have no credit history, informal addresses or few documents. Sarda said the answer is not to relax the rules.

“Don’t lower the bar. Widen what counts as evidence.”

He said checking local registries in both the sending and receiving country solves much of the problem, while small transfers can move with lighter checks under a risk-based system.

Stablecoins Move Fast But Leave A Trail

MoneyGram already supports stablecoin transfers. Sarda said the compliance basics for stablecoins are the same as for regular money, but the payments settle in seconds without banks screening them along the way. The upside is that the public ledger makes the money traceable.
Also Read: Crypto Firm ONDO Floods Robinhood Chain With 15 Stock And Fund Tokens
He named self-custody wallets, patchy enforcement of the Travel Rule and the businesses that convert crypto to cash as the biggest gaps today.

MoneyGram CEO Says Cash Still Matters

Apps like Wise and Remitly have pushed remittance fees down for years. MoneyGram CEO Anthony Soohoo told fathom.news he does not see its shops and its apps as rival networks.

“As money becomes more digital, access to the physical world becomes more valuable, not less.”

He said MoneyGram links digital wallets, bank accounts, cards and stablecoins to its retail locations, and is cutting costs through smarter routing and newer rails such as blockchain. In his view, digital money still needs a last mile, and MoneyGram is one of the few networks that has one.

The Real Risk Is A Patchwork Of Rules

Asked which regulatory change worries him most, Soohoo pointed to fragmentation rather than regulation itself. Cross-border payment firms already deal with hundreds of different rulebooks.

“The challenge is not really regulation. It’s more about fragmentation, but I think good regulation creates trust and common rules.”

He cited the US Genius Act, which sets rules for stablecoin reserves, redemption and supervision, as the kind of clear framework that gives MoneyGram confidence to move ahead. Where rules stay unclear, he said, companies tend to hold back.

A Crowded Market For ID Checks

Dozens of vendors now sell similar verification tools. Sarda said simple data lookups have become a commodity, but reliable onboarding decisions across 200+ countries have not. He expects single-country and single-check tools to get squeezed out as the market consolidates.
AiPrise is a Santa Clara, California, startup backed by Y Combinator. It offers business checks, identity checks, anti-money laundering screening and document verification through a single API, and counts payment firms, neobanks, marketplaces and crypto platforms among its customers.
Read Next: Invesco Hands Advisers 2 Claude Tools To Build Portfolios Faster

Originally published on Fathom News.

Invesco Hands Advisers 2 Claude Tools To Build Portfolios Faster

Fathom News — Wed, 28 Oct 2026 11:00:40 +0000

Investment giant Invesco is putting its portfolio rules and product data into Anthropic's Claude, an AI chatbot, so financial advisers can build client portfolios and handle paperwork faster.

Key Takeaways

Invesco is bringing custom portfolio and practice management capabilities into Anthropic’s Claude for financial advisers
The integration will put Invesco data and tools inside Claude so advisers can use natural language to support portfolio construction and business management tasks
Anthropic has been deploying industry-specific Claude implementations and focusing on financial adviser use cases via partner releases
Invesco supplies portfolio rules, product data and firm playbooks so Claude generates responses using the model’s reasoning plus the private material

The asset manager said the integration will put Invesco data and tools inside Claude so advisers can use natural language to support portfolio construction and business management tasks.

The firm is a publicly traded asset manager that sells investment products and adviser services, while Anthropic sells Claude as a general AI assistant and as an enterprise platform. Anthropic’s broader adviser push appears on its announcements page, where the company has been stacking industry-specific deployments rather than only selling generic chat access.

Invesco Claude Moves Beyond General Chat

The key development is that a large asset manager wants Claude tied to narrow, high-value workflows where outputs affect portfolios, compliance reviews and how advisers run their books.
Wealth management firms prioritize whether an assistant can retrieve the right client context, surface product rules and reduce manual switching between systems rather than open-ended conversation.

Invesco’s release highlights two workflows in particular, custom portfolios and practice management. Those areas are profitable because they sit close to revenue and client retention.
If AI tools save time there, firms can justify real deployment budgets.

From Model Access To Embedded Distribution

Over the past year many financial firms experimented with large language models behind internal walls, but projects often remained shallow because staff still had to leave the chat window and return to legacy systems. This phase moves the data, rules and actions into the assistant itself instead of asking advisers to paste data into a chatbot.
That makes the AI interface more useful but raises the cost of governance failures.

Also Read: Anthropic Slashes Claude Prices 40% For Government Buyers

As a result, firms such as Invesco are acting as distributors of domain context, not just buyers of model seats. The commercial value lies less in raw model intelligence and more in the proprietary workflow wrapped around it.

Claude’s Adviser Push Keeps Narrowing

Anthropic has been leaning into financial adviser use cases through multiple partner releases, suggesting it views wealth management as a tractable enterprise market where trust, document retrieval and guided recommendations matter more than flashy generation.
A generic assistant competes on price. A workflow assistant tied to a firm’s own content can command a higher budget because it saves labor within existing revenue lines.

The risk remains that finance produces confident-looking partial answers.
Advisers and compliance teams will need clear lines on what Claude can recommend, what it can only summarize, and where a human must sign off.

How Adviser AI Gets Grounded In Firm Data

In the Invesco setup, Claude receives fresh context at the moment of use, including Invesco Claude portfolio rules, product data and firm playbooks supplied by the asset manager. Claude then generates responses using both the model’s reasoning and the private material it was allowed to read for that request.

For registered investment advisers, consistent process often matters more than one perfect answer.
A tool that reduces searches across planning systems, documents and product libraries can save minutes on each client interaction that compound across a practice.

What changes next is whether firms stop treating AI as a sidecar. If Invesco and Anthropic can keep responses bounded to approved data, the assistant moves from chatbot toward an operating layer for adviser software.

Originally published on Fathom News.

India Telecom Giant Jio Soars 21% As $3.2B Debut Tests AI Plan

Fathom News — Wed, 28 Oct 2026 09:15:18 +0000

Jio raised about $3.2 billion in its IPO, and its stock jumped more than 21%, putting pressure on its AI and cloud plans to start making real money.

Key Takeaways

Jio Platforms entered public markets this week, raising about $3.2 billion in the Jio IPO
The public listing and filing created a disclosed valuation and debut market price for Jio Platforms
Akash Ambani used the listing to pitch AI, cloud and space connectivity as the company’s next growth engines
Reuters reported the stock surged more than 21% on debut

The company’s BSE filing documents the listing event that is driving fresh interest in the Jio IPO, even though some search chatter framed it as a draft filing with SEBI rather than a completed market debut. The official record now is a listed company setting out its next act, not a private company just starting the IPO process.

Jio IPO Creates A Public Valuation

The listing and the public filing make the arithmetic simple. Disclosure of the event, the raise of about $3.2 billion and the debut market price create a public valuation that must now be tested against operating performance.

AI Bet, The New Pitch Is Not Just Telecom

Akash Ambani used the listing to pitch AI, cloud and space connectivity as the company’s next growth engines, shifting attention from subscriber scale to computing services.
He said Jio’s next decade would focus on AI, cloud computing, connected devices and space-based connectivity, and that framing connects Jio’s network reach to an enterprise and developer-facing services strategy.
Jio already markets Jio Brain, a machine learning platform that trains models and runs them across cloud and edge infrastructure closer to users and devices. The official filing and debut gave public investors a clean event to price.
Reuters reported the stock surged more than 21% on debut, which turned a telecom success story into a live test of whether investors will back Jio’s broader AI and cloud narrative. Retail interest often starts with share price curiosity.
Also Read: SEC Puts 3 Firms On Public List For Tokenized Stock Trading

Can AI Drive Recurring Revenue

The harder question now is whether Jio can turn network reach into paid AI services at Indian scale and convert that into recurring revenue rather than one-off growth expectations. What has to be true for the new valuation to make sense is arithmetic, disclosed capital from the IPO plus any public-market implied market cap must be convertible into sustainable cash flow from AI, cloud or satellite services.
That requires measurable revenue lines attributable to AI or cloud in future filings, realistic margins after the capital-intensive build of data centers and edge infrastructure, and customer willingness to pay for services over and above existing connectivity.
Revenue and the expectation of revenue must be kept separate, a bullish narrative does not equal reported revenue today, and future filings will need to show the transition from expectation to actual operating results.

Execution And Capital Challenges

Jio’s challenge is execution rather than attention, AI infrastructure is capital intensive, cloud competition is crowded and satellite or space connectivity adds another long investment cycle. Still, the listing gives Jio both capital and a public scorecard.
If future filings start breaking out revenue from AI or cloud, Ambani’s AI Bet will move from promotional narrative to a quantifiable part of Jio’s operating story.
Read Next: AI Agents Can Go Off The Rails After 5 Or 50 Hidden Steps

Originally published on Fathom News.

AI Chip Startup Tenstorrent Bags $1.5B, Hits $6.2B Value

Fathom News — Wed, 28 Oct 2026 08:00:26 +0000

AI chip startup Tenstorrent said on Oct. 27 it raised $1.5 billion at a $6.2 billion value to build more chips and software tools, giving cloud companies another supplier beyond Nvidia.

Key Takeaways

Series E closed at $1.5 billion and valued Tenstorrent at $6.2 billion, the company announced Oct. 27
The funding round lists Innovation X as a co-lead
Tenstorrent builds processor cores and compiler stacks aimed at training and inference
The company plans phased rollouts, ramping foundry orders, expanding compiler and runtime teams, and fielding reference systems

The funding round lists Innovation X as a co-lead.

What Tenstorrent Said And The Deal's Scale

Tenstorrent says Innovation X co-led the round and that the funding will support the company's next phase of growth, including ramping foundry orders and expanding engineering teams. It added that the round valued the company at $6.2 billion.

Why This Matters To AI Infrastructure Buyers

The company builds processor cores and compiler stacks aimed at training and inference.
Its architecture competes directly with established GPU designs by offering a mix of scalar and matrix engines and a compiler that maps AI graphs to chiplets. If it ships at scale, cloud providers and AI labs could gain a second supply source beyond incumbent GPU vendors, easing procurement bottlenecks and providing negotiating leverage for large buyers.
The Series E funding therefore matters to cloud buyers, enterprise AI teams, and chipfoundries negotiating capacity.
Also Read: AI Chip Startup Infinity Grabs Meta Talent To Challenge Nvidia

How Tenstorrent Will Use The Money

The company plans phased rollouts, ramping its foundry orders, expanding compiler and runtime teams, and fielding reference systems for hyperscalers. The release says part of the capital will fund test deployments in enterprise data centers and developer tools to shorten integration time for AI frameworks.

What To Watch Next

Watch the startup for shipments to cloud partners, early performance claims against incumbent GPUs, and announcements about foundry partners and yields.
Quarterly filings or company updates that reveal unit economics, tape-out results, or customer commitments will be the clearest proof points that the Series E moves the company from R&D to production.
Read Next: AWS Case Study Says Exowatt Slashes AI Bills And Beats Big Rivals

Originally published on Fathom News.

AI Cloud Firm Together Saves Costly AI Jobs By Reviving Broken Servers

Fathom News — Wed, 28 Oct 2026 07:15:17 +0000

Together, a company that rents out powerful AI servers, rolled out GPU Auto Repair on Tuesday to spot broken machines and fix them faster so customers do not waste time and money.

Key Takeaways

Together made GPU Auto Repair generally available on Tuesday
The system watches nodes both inside customer virtual machines and on the host below them
It can respond to Xid faults, double-bit ECC errors, PCIe failures and unstable InfiniBand links
Customers can set repairs to wait for approval or run automatically and set escalation timing for active workloads

Together posted the update on X on Tuesday. The company rents and operates AI cloud infrastructure for model training and inference workloads.

In a detailed blog post, Together said the system watches nodes both inside customer virtual machines and on the host below them.
It can respond to Xid faults, double-bit ECC errors, PCIe failures and unstable InfiniBand links.

That matters because modern training runs spread work across many GPUs at once. One bad node can delay the whole job, not just the machine that failed, wasting coordination across the full cluster as well as paid hardware time.

Together said customers can choose whether each repair waits for approval or runs automatically.
They can also set how long the system waits for active workloads to finish before escalating to a stronger action.

GPU Auto Repair Favors Small Fixes First

The core design choice is simple. The feature does not jump straight to a reboot.
It tries the lightest repair that matches the fault, then steps up only if that action fails.

That departs from many cluster recovery tools that key off a single NotReady signal. Together said that signal can look identical whether a node is truly broken or just briefly unreachable, which can trigger needless restarts.

In practice, the system works like a fault ladder.
Health checks spot a problem, classify it, suggest the smallest viable action and keep evidence attached to the recommendation. A customer can preapprove some classes of fixes and hold others for human review.

That mix of automation and consent matters for AI infrastructure buyers, because large training jobs are sensitive to interruptions and compliance teams do not want opaque software making destructive changes by itself.

Also Read: U.S. Grabs 72% Of AI Chips And Servers As China Gets 10%

Earlier Node Remediation Set Up This Release

Together said it first introduced automatic node remediation earlier in 2026 and had already been using it on internal inference systems and with cluster customers.
Tuesday's write-up turned that internal capability into a clearer product story.

That timing fits a wider industry shift. Over the past year, cloud sellers focused heavily on adding more GPUs.
Now they are spending more time on keeping those GPUs alive, visible and recoverable.

For buyers, that is a different kind of performance pitch. Raw chip count wins headlines, but uptime wins contracts when a weeklong run can burn through thousands of synchronized accelerators.

Where It Could Matter Most

The capability matters most for teams that run distributed training, where one failed worker can leave the rest idling, and for inference fleets that need to keep replicas healthy without bouncing whole services.
Together's system appears built around fault-specific policies.

A flaky network link might justify one action, while a repeated ECC memory fault might justify a stronger one. That is more useful than one blanket rule because GPU failures are not all the same.

The next test is coverage.
Together said preview support is coming for node-label remediation contracts that let outside detection systems feed the same pipeline. If customers plug in their own monitors, the feature stops being a single cloud capability and becomes part of a larger operating system for AI clusters.

Why Fault Types Shape Repair Paths

Together named several faults that trigger the system, including Xid errors, ECC errors and InfiniBand link failures.
Those are not interchangeable problems, since Xid errors often point to GPU driver or hardware issues, while InfiniBand failures hit the network fabric that ties many GPUs together.

That distinction matters for operators buying cluster time from Together. A reboot might help one fault and worsen another by killing an active training run.
The tool tries to map each symptom to a narrower response before escalating.

For model builders, the practical gain is shorter mean time to recovery. For finance teams, the gain is simpler to count, because every avoided restart saves paid GPU minutes that would otherwise vanish while a large distributed job waits for one unhealthy node.

Originally published on Fathom News.

Anthropic Slashes Claude Prices 40% For Government Buyers

Fathom News — Wed, 28 Oct 2026 04:30:23 +0000

Anthropic cut Claude For Government prices by 40% for public agencies through a new federal buying deal, aiming to save taxpayers roughly $1 billion.

Key Takeaways

Anthropic cut Claude For Government prices by 40% for federal, state, local, tribal and territorial buyers under a new GSA OneGov agreement
The OneGov deal extends discounted access to products used by about 500 government organizations
The discount now applies across Claude Enterprise, Claude For Government and the Claude API, and agencies can buy directly now
Utah gave Claude Code to 34 developers across 10 state agencies, with 80% reaching a working environment within 30 minutes

Anthropic posted the new OneGov agreement on Tuesday evening. The deal extends discounted access to products used by about 500 government organizations.
The company sells large language model tools to businesses and governments through chat products, coding tools and application programming interfaces.

Anthropic also said the discount now applies across Claude Enterprise, Claude For Government and the Claude API. The company said agencies can buy directly now and through cloud providers in coming months.

That breadth matters because Claude For Government had been a narrower procurement path before this agreement.
Agencies often buy one surface at a time, which slows pilots, limits scaling and forces separate approvals for chat access, developer use and backend integrations.

Anthropic gave a few adoption figures that show why it pushed for a wider deal. Weekly Claude For Government active users are up about eightfold since February, and use across federal civilian agencies has grown about 20-fold over the past year.

The company also tied the deal to tight public budgets and long procurement cycles.
That is a blunt sales point, but it fits how government software buying works. A 40% cut can matter more than a new model benchmark when agency buyers need to justify each contract line.

Claude For Government Spreads Beyond Pilots

Claude For Government is Anthropic's public sector version of Claude.
It is built for agencies that need tighter controls around data handling, access rules and deployment paths than a consumer chatbot offers.

Also Read: Anthropic’s Claude Joins 6 Drug Firms Hunting A Cure For Rare Muscle Disease

Anthropic said Utah gave Claude Code to 34 developers across 10 state agencies, with 80% reaching a working environment within 30 minutes. It also said 71% reported saving more than 20 hours by the end of a four-week pilot.

Those examples point to a practical sales pitch.
Anthropic is not selling abstract AI transformation here. It is selling backlog reduction, code modernization and incident response speed to offices that still run old systems.

From Federal Beachhead To OneGov Deal

The new agreement follows a broader push by AI labs into regulated and public sector markets.
Earlier in October, large enterprise AI vendors kept adding controls for identity, billing and security as buyers demanded more than chatbot access.

Anthropic had already been building that case. Two hours before this item, Fathom covered a separate Anthropic tie-up focused on rare disease work, showing how the company is trying to turn Claude into a domain product, not just a general assistant.

For public agencies, that shift matters because procurement no longer centers only on model quality.
It centers on whether a tool can fit existing workflows, contract vehicles and oversight rules.

How GSA Buying Changes AI Access

The GSA's OneGov structure acts like a centralized buying path. Instead of each office negotiating a fresh deal, agencies can use a common framework that shortens paperwork and standardizes terms.

For Anthropic, that means Claude For Government can move from isolated pilots into repeatable deployments.
For state and local buyers, it can lower the gap between "we tested this" and "we budgeted for this."

The next thing to watch is whether agencies expand from chat seats into API usage. Claude Enterprise can help staff draft and search, but Claude API spending is where workflow automation starts.
If that mix changes over the next few months, the real winner will not be the chatbot budget. It will be the legacy software contractors Claude begins to replace.

Originally published on Fathom News.

Oracle Hands Google AI Chips To Cloud Buyers Eyeing Nvidia

Fathom News — Wed, 28 Oct 2026 02:15:18 +0000

Oracle, the business software and cloud company, will add Google TPUs to its cloud so big companies can build AI without relying only on Nvidia chips.

Key Takeaways

Oracle will offer Google's tensor processing units inside Oracle Cloud Infrastructure
The move gives enterprise buyers another way to train and run AI systems without relying only on Nvidia-based fleets
Oracle announced the TPU move during its AI World event, where it also rolled out broader AI products tied to OpenAI and Snowflake
Oracle still needs to show which Google TPUs arrive first, in which regions, on what pricing and with which software support

Oracle announced the move during its AI World event, where it also rolled out a broader set of AI products tied to OpenAI, Snowflake and new security tools. The Register reported further details on how Oracle plans to position the offering for customers waiting for clearer returns on AI spending.

The immediate point is choice.
Most enterprise AI demand has piled into Nvidia systems because they are widely supported by model frameworks and cloud tooling. Google TPUs offer a different accelerator design, built for machine learning workloads, and Oracle is betting that some customers want that option inside OCI rather than only inside Google Cloud.

Oracle Cloud Infrastructure is Oracle's public cloud platform for compute, storage, databases and AI services.
OCI has spent the past year trying to win buyers that want high-performance AI clusters but do not want all of their data or applications tied to one cloud vendor.

Google TPUs Give Oracle A New Sales Angle

For Oracle, the pitch is less about beating Google on silicon and more about absorbing more AI demand onto OCI. If a customer prefers Google TPUs for model training but wants Oracle databases, enterprise software or regional cloud capacity, Oracle can now try to keep that workload in-house.

That could matter for corporations building mixed stacks.
A bank may want one model optimized for Google TPUs, another tuned for Nvidia GPUs and customer data stored close to existing Oracle systems. Before, that often meant crossing clouds and accepting more operational friction.

Google has long treated TPUs as a differentiator for its own cloud.
Putting Google TPUs into OCI suggests the competitive boundary is shifting from exclusive hardware access toward broader distribution, as long as the commercial terms still work for both sides.

Also Read: SpaceX Plows $40B Into Nvidia Chips For AI Cloud

Oracle's AI Week Keeps Getting Broader

Tuesday's disclosure follows a burst of Oracle AI product news across infrastructure, applications and security. Earlier disclosures focused on Oracle's use of Nvidia racks and deeper links with OpenAI.
The TPU step broadens that message from software integration toward hardware optionality.

That matters because cloud AI buyers now ask two separate questions. First, can a provider get enough chips.
Second, can it offer more than one hardware path when costs, latency or model compatibility change. Oracle is trying to answer yes to both.

The strategy also speaks to a larger market shift.
Clouds no longer win AI deals only by owning the most chips. They also need to connect those chips to customer data, compliance controls and existing software budgets.

Also Read: Nvidia Turns To Synopsys To Catch Costly Chip Errors Before Build

What Buyers Should Watch Next

The next test is execution.
Oracle still needs to show which Google TPUs arrive first, in which regions, on what pricing and with which software support.

Those details decide whether Google TPUs on OCI become a niche option or a real second lane for enterprise AI. If Oracle can make migration and orchestration simple, it strengthens its hand in a market still dominated by a few chip and cloud suppliers.

How Google TPUs Fit Into OCI

Google TPUs are custom chips built for tensor operations, the matrix math that powers model training and inference.
Nvidia GPUs began as graphics processors and later became the default AI workhorse because developers could program them flexibly. TPUs started from the opposite direction. Google built them to run machine learning efficiently at scale inside its own infrastructure.

For Oracle customers, the practical change is not the chip brand alone.
It is the ability to keep applications, databases and governance inside OCI while selecting a different accelerator under the hood. A retailer using Oracle databases could train a recommendation model on Google TPUs, then deploy it close to its transaction systems without shifting the whole workload into Google Cloud.

That matters most for enterprises with old software estates.
Many large companies already depend on Oracle for finance, HR or databases. If Oracle can present Google TPUs as another OCI resource, those firms get more flexibility without rewriting every surrounding system.

Originally published on Fathom News.

AWS Case Study Says Exowatt Slashes AI Bills And Beats Big Rivals

Fathom News — Wed, 28 Oct 2026 00:30:32 +0000

Exowatt, a software company, used SageMaker Tuning to build cheaper custom tools that AWS said were up to 49x less costly and more accurate than bigger rivals.

Key Takeaways

Exowatt used SageMaker Tuning to build two task-specific AI models that beat frontier models on accuracy while cutting costs by up to 49x
AWS said Exowatt used Amazon SageMaker AI serverless model customization to create a repeatable pipeline for edge and cloud deployments
An AWS post explained evaluation bias can cause teams to overestimate large general models without clean ground-truth tests
SageMaker Tuning was serverless so Exowatt did not have to reserve dedicated tuning infrastructure before each job

AWS said Exowatt used Amazon SageMaker AI serverless model customization to create a repeatable pipeline for both edge and cloud deployments, and a second AWS post the same day on evaluation bias explains why companies can overestimate large general models without clean ground-truth tests. Exowatt is building AI for operational use cases where cost and latency matter as much as model breadth.

SageMaker is Amazon's managed platform for training, tuning and deploying models without forcing customers to build the infrastructure stack from scratch.
The claim is not that frontier models stopped mattering, but that many enterprise tasks need smaller models tuned on the right data, tested on the right benchmark and cheap enough to run at production scale.

SageMaker Tuning Beats Bigger Models On Narrow Tasks

This distinction has become a major shift in business AI, with frontier models performing broadly but extra generality adding cost, latency and operational complexity for firms that need a narrow function. AWS said Exowatt built a pipeline that produced specialized models for two deployment targets, important because edge systems often require smaller footprints and more predictable response times than cloud-only applications.

The 49x figure is the part buyers will remember, because cost gaps of that size can decide whether a feature stays in pilot or becomes a shipped product.
If a custom model also beats the larger baseline on accuracy, the argument for defaulting to the biggest available model weakens.

Why Clean Evaluation Suddenly Matters More

The companion AWS discussion of evaluation bias addresses a common AI buying mistake, where teams often compare a new model against labels produced by an incumbent model and treat the result as objective. That can hide gains from a specialized alternative.
If the benchmark inherits the older system's mistakes, a newer model that improves real outcomes may still look worse on paper.

Exowatt's result only matters if the evaluation actually measures the task it wants to solve. This also explains why many enterprise AI case studies now focus on workflow metrics instead of leaderboard bragging.
What matters is whether a model completes a business task faster, cheaper and with fewer errors.

Also Read: Agentic LLM Inference Warning Calls Labs Halve Latency Cut Costs 83%

What SageMaker Tuning Means For Model Buyers

The next watchpoint is whether AWS can turn SageMaker Tuning into a repeatable enterprise habit rather than a few strong case studies. Many companies still buy model access via a shared API and a broad prompt, fine for experiments, expensive in production.
If SageMaker Tuning keeps producing narrower systems with better economics, enterprise AI may keep moving from one-size-fits-all models toward fleets of smaller customized ones.

How SageMaker Tuning Cuts The Bill

SageMaker Tuning adapts a base model to a narrow job with specific examples, labels or prompts, then packages it for a target environment. AWS said Exowatt’s process was serverless, so it did not have to reserve dedicated tuning infrastructure before each job, lowering setup friction and shortening iteration cycles.

For Exowatt, the gain comes from matching model size to task size.
A large frontier model carries broad knowledge, much of which may be wasted on a limited workflow, while a smaller tuned model concentrates compute on the domain it faces, allowing lower cost and higher accuracy to coexist.

For Amazon, cases like this support a broader pitch, with AWS wanting companies to build optimized model stacks inside SageMaker and Bedrock, expanding cloud revenue beyond inference.

Originally published on Fathom News.

U.S. Grabs 72% Of AI Chips And Servers As China Gets 10%

Fathom News — Tue, 27 Oct 2026 22:49:10 +0000

Researchers said the U.S. has taken 72% of the world's AI compute since 2022 while China got 10%, giving America a big edge in building AI tools people and companies use.

Key Takeaways

Epoch AI Research said 72% of the world's AI compute since 2022 ended up in the U.S., while China accounted for about 10%
The U.S. captured about 60% of new AI compute added each year from 2022 through 2025, Epoch said
Epoch did not publish the full methodology in its X post, so the figures should be read as directional estimates
Oracle, Amazon and Microsoft spend heavily on whole systems, not just processors, to support large model training and inference

Epoch AI Research posted that the U.S. also captured about 60% of new AI compute added each year from 2022 through 2025. The research group studies long-run trends in model training, hardware demand and AI infrastructure.

The figures matter because AI compute, the chips and servers used to train and run models, now shapes who can build frontier systems at scale.
A lab can improve data and algorithms, but it still needs thousands of accelerators, power contracts and networking gear to move from prototype to production.

Epoch did not publish the full methodology in the X post, so the figures should be read as a directional estimate rather than a census of every GPU in service. Even so, the split lines up with recent industry moves, including large cloud buildouts in the U.S. and export controls that have slowed China's access to leading-edge accelerators.

AI Compute Concentrates In US Clouds

The concentration also helps explain why so many model developers have tied themselves to American infrastructure partners.
When most AI compute sits inside U.S. clouds and colocation campuses, access becomes a commercial negotiation before it becomes a research question.

That dynamic has widened the gap between labs that can secure long-term capacity and those that cannot. It also pushes countries outside the U.S. to subsidize domestic compute, court foreign chipmakers or settle for older hardware that raises training costs and lengthens deployment cycles.

For cryptocurrency readers, the pattern is familiar.
Scarce AI compute now behaves a bit like scarce block space or scarce mining capacity. Whoever controls the supply can charge for access, prioritize favored workloads and shape the economics of everyone downstream.

Also Read: Drug Research Startup Bags $215M To Build U.S. Chemistry Labs

Four Years Of Policy And Procurement

The imbalance did not appear overnight.
Since 2022, U.S. hyperscalers have signed multiyear chip commitments, leased new power at industrial scale and absorbed much of the world's top-end accelerator output.

China, by contrast, has had to work around export rules, domestic chip limits and tighter access to advanced packaging and memory. That does not stop Chinese firms from building strong models, but it changes the cost curve and slows the rate at which new AI compute can come online.

The latest figure also lands after a year in which investors have treated compute access as a proxy for AI durability.
Revenue claims matter less when the harder question is whether a company can still reserve enough machines six months from now.

What AI Compute Changes For Everyone Else

The next thing to watch is whether the U.S. share keeps rising even as other regions spend more. If Europe, the Gulf and parts of Asia add meaningful capacity in 2027, the market may become less concentrated without becoming truly balanced.

For startups, the practical issue is simpler.
If AI compute keeps clustering in a few regions and clouds, model makers may have to choose between higher margins, lower latency or strategic independence. Few will get all three.

How AI Compute Turns Into Power

Epoch AI Research's figures describe more than chip counts.
In practice, AI compute means the usable training and inference capacity inside clusters built from accelerators, networking switches, memory and power delivery. A company with 10,000 advanced GPUs but weak interconnects may have less effective AI compute than a rival with fewer chips and faster networking.

That is why Oracle, Amazon and Microsoft spend heavily on whole systems, not just processors.
Training large models requires data to move across racks with low delay, while inference requires enough available servers to answer user requests at steady cost. When Epoch says the U.S. won 72% of AI compute, it points to control over the full stack, land, electricity, cooling, chips and cloud software.

For Chinese labs and cloud firms, the number matters because each missing layer compounds the next constraint.
A shortfall in top-end chips can force smaller clusters. Smaller clusters can mean longer training runs.
Longer runs can delay product cycles and raise financing needs. That is the chain reaction behind one statistic.

Originally published on Fathom News.

J.P. Morgan Puts Its Tax-Cutting Tool Inside Claude For Financial Advisors

Fathom News — Tue, 27 Oct 2026 22:47:52 +0000

J.P. Morgan's 55ip unit is plugging its tax-cutting investment software into Anthropic's Claude, so financial advisors can rebalance client portfolios and trim tax bills without leaving the chat.

Key Takeaways

55ip, part of J.P. Morgan Asset Management, is adding its Tax Smart tools to Claude for Financial Advisors
The tools help advisors build portfolios and choose trades that cut a client's tax bill
Investment tech firm Conquest announced its own Claude for Financial Advisors tool the same day
Neither company shared prices, deal terms or how many advisors will use it

55ip said on Tuesday that its software will work inside Claude for Financial Advisors, Anthropic's version of its AI assistant built for people who manage other people's money.
55ip makes software that helps advisors put together portfolios and make trades with taxes in mind. Advisors will now be able to use those tools by asking Claude, instead of switching to a separate program.

How The Claude For Financial Advisors Tax Tool Works

When an advisor rebalances a client's investments, some sales can trigger a tax bill. Tax-smart software picks which holdings to sell, and when, to keep that bill as small as possible.
Inside Claude, the advisor can ask for that work in plain words. Claude passes the request to 55ip's own system, which holds the client's account details and does the tax math, then shows the result.
Also Read: HCLTech Becomes Anthropic's Top Claude Partner, Builds New AI Business Unit
That matters because an AI assistant on its own does not know what a client owns or what they paid for it. The link to 55ip gives Claude that information in a way the advisor's firm controls.

More Firms Join Claude For Financial Advisors

55ip is not alone. Investment tech firm Conquest announced on the same day that its financial advice tool is also coming to Claude for Financial Advisors.
The launches came a day after Anthropic filed to go public, a moment when investors are watching closely for signs that Claude can earn steady money from businesses.

What We Still Do Not Know

55ip did not say what the service costs, when every advisor will get it or how many firms have signed up. The next test is whether big advisory firms move their daily work onto Claude, or only try it out.
Read Next: Oracle Hands Partners $1B To Cut AI Cloud Bills For Big Firms

Originally published on Fathom News.

Small Investor Blockmate Bets On Nvidia Server To Cash In On AI

Fathom News — Tue, 27 Oct 2026 22:31:15 +0000

Blockmate Ventures, a small public investment company, bought one Nvidia B200 Node to start renting computing power and tap rising demand for AI tools.

Key Takeaways

Blockmate Ventures acquired one Nvidia B200 Node on Tuesday
The company said the purchase marks a shift from venture building to owning and operating AI compute
The announcement did not disclose the purchase price, expected utilization, customer pipeline or hosting location
A B200 system is built for training and inference on large AI models, and operators rent that capacity by the hour

Blockmate said the purchase marks a shift from venture building to owning and operating AI compute, while Nvidia has separately said it is expanding supply chain capacity to speed AI infrastructure buildouts. The announcement did not disclose the purchase price, expected utilization, customer pipeline or hosting location, and that matters because a single Nvidia B200 Node is not a hyperscale build but a beachhead.

Blockmate AI Cloud Starts Small

A B200 system is built for training and inference on large AI models, and operators rent that capacity to startups and enterprises by the hour or under reserved contracts.
For smaller public companies, even one node can open a higher‑margin business line than software services or passive venture exposure, but idle GPUs destroy the economics quickly because depreciation starts on day one.

Blockmate framed the deal as its first owned compute asset, and ownership matters operationally, unlike brokers, consultants or shell partners, a firm that owns the node controls the asset that produces billable GPU time and has committed capital that indicates intent beyond mere marketing claims.

From Venture Builder To Owned Compute

The move follows a broader scramble across AI infrastructure markets since late 2025, when demand for advanced Nvidia systems began outrunning available supply across both cloud and enterprise channels. Smaller entrants have tried to carve out room beneath hyperscalers by buying limited clusters and reselling capacity to labs that cannot secure large contracts.

That strategy carries clear execution risk because a neocloud operator needs power, cooling, interconnects, scheduling software and enough customer demand to keep the machine full.
If a B200 node runs near full occupancy, revenue compounds fast since the asset is already installed, but if it sits idle financing, colocation and depreciation will quickly erode margin, so the next useful numbers will be utilization, contract length and revenue per GPU hour.

In practice, physical constraints, power availability, datacenter rack capacity and interconnect bandwidth, often determine growth more than sales pitches do. Capital spending speaks louder than strategy slides, and a company that commits cash to buying scarce machines is signaling operational intent.
Investors should treat that spending as a stronger indicator of future capacity than marketing claims.

Also Read: AMD's MI355X Takes On Nvidia In First Head-To-Head AI Agent Test

What Nvidia B200 Node Can Change

A single node will not alter Nvidia supply or cloud pricing, yet it can still matter for a company Blockmate's size if the machine becomes the first proof that it can turn AI demand into recurring infrastructure revenue. Investors should watch deployment timing, whether Blockmate publishes customer names or reservation data, and whether it buys more hardware or signs hosting agreements that move it beyond a one‑node test.

The unit economics hinge on utilization and the cost per unit of inference.
Power and cooling increase marginal operating cost, while financing terms and depreciation set the baseline cost that must be covered by revenue per GPU hour. If Blockmate can maintain sustained high utilization, marginal margins improve rapidly because the fixed capital is already in place.
If not, the asset will be a capital drain.

How Nvidia B200 Supply Shapes Small Clouds

Nvidia's B200 systems sit near the top of the current AI hardware stack because they combine fast memory, dense interconnects and software support that many model developers already use, so access is often the binding constraint rather than sales capability for a company like Blockmate Ventures.

A neocloud works by buying scarce machines up front and renting slices of them to customers that cannot get capacity from Amazon, Microsoft or CoreWeave on acceptable terms, and the business succeeds or fails on sustained utilization and the resulting cost per unit of inference.

For smaller public firms, one node is also a signaling device that shows investors and customers the company owns real compute, not just a memorandum of understanding or reseller pitch, and the next test for Blockmate Ventures is whether that signal turns into disclosed contracts.

Originally published on Fathom News.

Free AI Reader Turns Messy Scans Into Text And Flags Its Own Mistakes

Fathom News — Tue, 27 Oct 2026 22:30:34 +0000

Open-source AI lab OpenBMB has released MiniCPM OCR, a free tool that turns scans, phone photos and PDFs into text and marks the spots it may have misread.

Key Takeaways

OpenBMB released MiniCPM OCR, a 1.3 billion-parameter OCR model family for PDFs, scans and photographed documents
The GitHub repository says the main model scored 93.68 on Real5-OmniDocBench and improved end-to-end processing efficiency by 55%
The repository says the system handles Chinese and English documents, tables, formulas, charts and key information extraction
The GitHub page says the model runs about 1.49 times faster than PaddleOCR-VL-1.6 and 3.62 times faster than MinerU2.5 Pro

OpenBMB posted the launch on X and linked a GitHub repository describing MiniCPM OCR 1.0, a suite for text recognition, document parsing and error review. The repo says the main model scored 93.68 on Real5-OmniDocBench and improved end-to-end processing efficiency by 55%.

OpenBMB builds open-source multimodal models that mix text and vision. OCR, short for optical character recognition, converts text in images or scanned files into machine-readable content.

MiniCPM OCR Targets The Boring Work AI Still Misses

Document parsing remains one of AI’s most expensive failure points.

Models can answer hard questions in chat yet still miss a seal, a table cell or a number on a warped invoice. The repository says the system handles Chinese and English documents, tables, formulas, charts and key information extraction.

It also includes an evaluation component that flags suspect regions rather than treating every output as equally trustworthy.

That region-level confidence is crucial in financial, legal and health workflows where users need to know which part of a page still requires human review, not just that text was extracted.

Also Read: Semantic Scholar Lab Unleashes Free Tool That Finds Papers Searches Miss

The New Release Follows A Broader OpenBMB Push

The launch is separate from OpenBMB research released Monday about human nudge performance in agentic tasks, which examined model decision making. This release is a product move into document ingestion, a different layer of the stack.

The timing follows a wider industry pattern, with multimodal labs taking on front-end tasks once left to specialist software, from coding and support to speech and paperwork.

OCR feeds retrieval systems many AI agents depend on, if a page is read badly, downstream steps inherit mistakes, which is why even strong large language models can fail on messy real-world files.

Where The Speed Claims Matter Most

The GitHub page says the model runs about 1.49 times faster than PaddleOCR-VL-1.6 and 3.62 times faster than MinerU2.5 Pro under the listed decoding setup. Those comparisons matter where thousands of pages hit a queue every hour.

Faster processing determines whether a bank, insurer or business process outsourcer can review documents in near real time or needs overnight batches and exception teams.

Benchmarks are useful, but production OCR often breaks on ugly inputs, phone photos, curved pages and low-contrast scans. OpenBMB says the system was trained for those cases.

Real-world users will test whether those claims hold outside curated datasets.

How Confidence Scoring Changes OCR Review

The repository describes MiniCPM-OCR Jev, which evaluates results at the region level and marks areas that need review. Traditional tools often return one final text output and force staff to hunt for mistakes line by line.

A region-level checker narrows the search by pointing reviewers to the exact table, stamp or paragraph likely to be wrong.

For operations teams, that can cut review cost without pretending the model is perfect. A bank processing mortgage files benefits from something that parses most pages quickly and highlights the small share of risky fields before they hit underwriting or compliance systems.

Originally published on Fathom News.