129 articles tagged
AWS made AgentCore payments generally available on 18 August 2026 in 12 Regions. The pricing page says no additional charge, then bills an API that does not exist and omits both gas fees and a mandatory subscription.
AWS gave Aurora DSQL a CloudWatch Database Insights metric on 20 August 2026. The mode model does not list DSQL and the AWS Price List API carries no Aurora DSQL Database Insights SKU.
AWS Lambda MicroVMs went from 5 regions to 10 on 19 August 2026. The AWS pricing API shows Frankfurt and Singapore running about 26% above N. Virginia, and six of the ten regions sit on a 400 GB account quota.
AWS added a task-hours cap to Security Agent on 19 August 2026. The smallest allowed value is 20 hours, and the Price List API rates pen testing at $50.00 per hour, so the floor is $1,000 a test.
AWS shipped five read-only Managed Dashboards on 14 August 2026 at no additional cost. You cannot schedule one directly, and the duplicate you need runs Cost Explorer API calls billed at $0.01 per request.
Bedrock now attributes bedrock-mantle cost by IAM user or role. The user guide says the finest grain is per usage type per day, tags are not retroactive, and existing CUR 2.0 exports must be recreated.
AWS shipped log group tag propagation for CloudWatch Centralization on 19 August 2026. The opt-in carries a rename requirement, a non-adjustable 5 TPS tagging ceiling, and a health field that stays green.
AWS says the three new CloudWatch pipelines processors cost nothing extra. The pipelines documentation says S3 and third-party sources are metered after processing, so every enriched byte is billed as a Custom log.
Amazon EKS now lets you tune the scheduler, controller manager and API server. Four parameters shipped on 12 August 2026, and the headline autoscaling one needs a control plane tier costing $1.65 an hour.
AWS said no VPC changes were required. The OpenSearch guide says every semantic enrichment index call on a VPC domain returns AccessDeniedException until you run AuthorizeVpcEndpointAccess first.
Vercel started metering Deployment Storage and Functions Storage at $0.10 per GB-month on 21 August 2026. The Pro retention default keeps production deployments for one year, and nobody set it.
The AWS announcement lists three exclusions. The admin guide lists five, plus an automatic Credential Guard shutdown and an AutoStop hibernation failure that costs real money to work around.
AWS made a second Las Vegas Local Zone, us-west-2-las-2a, generally available on 20 August 2026. Its own services table, updated 7 August 2026, omits the zone and contradicts the announcement on EC2 and EBS.
Anthropic shipped session budgets for Claude Managed Agents on 7 August 2026. The cap is priced at public list rates, overshoots by one model request per thread, and on a deployment it bounds every run separately.
Cursor's 19 August 2026 release runs each cloud subagent on its own VM. No machine-hour fee exists, so a five-way swarm costs roughly five times the tokens instead, and the docs restrict CI autofix to Teams.
DeepSeek's V4 rate card now has two prices for every token. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, off-peak is half, and one of the two peak blocks lands inside the Indian working day.
GitLab 19.3 removed the budget_caps_graphql_api feature flag and made GitLab Credits usage caps generally available. The release note says features are suspended at the cap. The subscriptions documentation says
xAI announced Grok 4.6 on Amazon Bedrock at $2 input and $6 output per million tokens. AWS's own model card prices two of the three inference paths at $2.20 and $6.60, and publishes no long-context tier at all.
OpenAI released a Prompt Caching dashboard on 20 August 2026 that reports cache reads per write. On GPT-5.6 that ratio decides whether caching cuts your input bill by 78 percent or raises it by 25 percent, and Azure
AWS says there is no additional cost to generate a SageMaker inference recommendation. Its own EAGLE documentation shows optimization jobs running on training instances for up to 120 hours across ten nodes.
Supabase relocated read replica management to Project Settings then Infrastructure on 21 August 2026. The interesting part is the cost model underneath it, and a live 10x pricing contradiction.
AWS announced customer managed key support for Timestream for InfluxDB on 20 August 2026. The developer guide hands you a --kms-key-id flag that six AWS reference pages do not document.
Vercel's new vcr command group landed on 20 August 2026. Two live Vercel pages disagree on repository limits, and the 500 MB layer ceiling stops fat images before the 15 GB total does.
Zed's 1.16.1 release on 19 August 2026 switched the default model for OpenAI subscription users to GPT-5.6 Sol. Its hosted price table still carries OpenAI's pre-30-July rates for Terra and Luna, so the published Luna
AWS made Web Search on Amazon Bedrock generally available on 4 August 2026 and told buyers to check the pricing page for costs. The pricing page has no price. Worse, the default IAM policy grants search but not external
AWS m6i.xlarge and Azure D4s v5 both cost $0.2020/hour in India as of 17 August 2026. Compute parity is not where migration budgets fail. This guide prices the replication meters, egress tiers, licence cliffs
Most data platform budgets miss because the build is quoted and the run is not. Here are the August 2026 list prices for Snowflake, BigQuery, Redshift, Databricks, dbt and Fivetran, what drives the bill, and how eCorpIT
Buying DevOps services in 2026 is mostly a decision about who owns the upgrade treadmill. Kubernetes ships a minor version every four months with 14 months of support, and AWS, Google and Microsoft all charge
Runtime instances bill EC2 plus a 12% management fee, and 7.8% on G-series. Worked against AWS's published microVM rates, the crossover on an 8 vCPU shape lands near 24% sustained utilisation.
Vector search in DynamoDB went generally available on 5 August 2026. Three billing meters, a 1 KB minimum, on-demand only, and a zero idle floor that reframes the pgvector and OpenSearch comparison.
Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 60% lower headline pricing. Three billing rules decide whether that discount survives contact with a real agent workload.
OpenAI cut GPT-5.6 Luna by 80% and Terra by 20% on 30 July 2026, and Amazon Bedrock passed both through automatically. Luna now costs $0.20 per million input tokens on both paths, which removes the pricing argument
Google Cloud Billing added an Originating products filter on 7 August 2026, so Gemini Enterprise spend no longer hides inside Vertex AI Search. The catch: the dimension is console-only, disables forecasting, and has not
GitHub's promotional AI credit allowance ends after 31 August 2026, cutting the pooled monthly allowance 37% for Copilot Business and 44% for Enterprise. Here is the arithmetic and the four defaults behind it.
Most observability overspend is not a vendor problem. It is a data-model problem meeting a billing rule nobody read. Here are the six published rules that move the number, with the primary sources.
GitHub shipped two Copilot cloud agent changes on 3 August 2026: a per-task reasoning level and automations that fire on issue or pull request comments. Both raise token spend, and GitHub's billing doc says nothing
PostgreSQL 14 gets its last community patch on 12 November 2026, and on Amazon RDS the paid clock starts 1 March 2027 at a published $0.100 per vCPU-hour. The real question is not whether to upgrade but which version
Gartner puts typical VMware licensing increases at 300 to 400 percent, and 48 percent of customers plan to shrink their footprint by 2028. Here is the cost model to run before you sign.
Published benchmark data, Apple's own specifications and one hosted price list are enough to work out when buying a Mac beats renting tokens. The answer depends on prefill, not generation, and most of the M5 numbers
Amazon charges $7 per 1,000 web search queries, Microsoft $14 per 1,000 transactions and Google up to $35 per 1,000 grounded prompts. The same 200,000-query month costs $1,000 to $5,425.
Only 20% of organisations could explain an overnight doubling of their AI bill within hours. This is the provider-by-provider attribution layer, the ownership model and the 90-day sequence that closes the gap.
Token prices moved in both directions inside five weeks in 2026. Here is how to meter an AI feature, pick a billable unit, run the margin arithmetic on real published rates, and cap the user who burns 100x the average.
Querying data that lives in another cloud can cost more in transfer fees than in compute. Here is how BigQuery Omni, Snowflake and Databricks each handle cross-cloud reads in 2026, with the per-TB numbers that decide
Datadog's Agent Observability bills on one thing: the count of LLM spans. Tool, retrieval and agent spans are free. Here is how the meter works, where the published pricing stops, and six ways to cap it.
Rapid Bucket is 5.5x the per-GiB price of standard Cloud Storage and bills for reads a same-region bucket serves free. On 50 TiB of checkpoints that is about $5,243 a month extra. Here is the arithmetic.
Notion's hosted runtime leaves free beta on 11 August 2026 and starts drawing Notion credits at about $0.0023 per run. Here is the arithmetic, the volume at which the bill overtakes a $5 Cloudflare Workers subscription
Gartner puts $234B of enterprise SaaS spend at risk from agentic arbitrage by 2030, and 97% of SaaS CEOs plan to drop seat pricing. Here is a practical framework to audit your stack, score each tool's exposure
AWS's remaining performance obligations jumped to $496B in Q2 2026 and much of 2027 capacity is already reserved. Here is what the backlog means for your cloud bill and a sourced FinOps playbook for choosing Savings
On July 1, 2026 Azure ended Reserved VM Instance purchases and renewals for 18 legacy VM series. Existing reservations run to term, then revert to pay-as-you-go, erasing up to 72% savings. Your options: savings plan
Enterprises run up to 660 SaaS apps, 53% of licenses sit idle, and Gartner puts $234B of SaaS spend at risk from agentic arbitrage by 2030. Here is how eCorpIT audits your stack, scores what to consolidate, and replaces
On August 1, 2026, Google Cloud began billing Semantic Governance Policies on the Gemini Enterprise Agent Platform, charging evaluation-model tokens for every governed agent action. Here are the metered rates, a worked
OpenAI's July 30, 2026 update drops GPT-5.6 Luna to $0.20/$1.20 and Terra to $2/$12 per million tokens and folds Priority Processing into a new Fast mode. What changes for your LLM spend, and how to migrate.
In 2026 an AI-driven memory shortage pushed DDR5 and server DRAM to record prices and stretched lead times. Here is what changed, why, and a sourced framework for CTOs choosing to buy hardware now, shift to cloud
In 2026, 98% of FinOps teams manage AI spend, and LLM prices span more than 140x from budget to frontier. An AI gateway with cost-aware routing and caching cuts that bill 40-70% without downgrading quality. How eCorpIT
Terraform moved to the BUSL license in 2023 and IBM now owns HashiCorp. OpenTofu, the MPL 2.0 fork under the Linux Foundation, has diverged. A 2026 decision guide: pricing, feature gaps and migration steps.
Five AI gateways, three real cost levers, and one decision. This 2026 comparison of LiteLLM, Cloudflare AI Gateway, Kong, Bifrost and OpenRouter covers routing, semantic caching, budgets, and where each one should run
Google's first-party Data Engineering Agent reached general availability on 22 April 2026. This decision guide compares it with dbt and Dataform on capability, cost, lock-in, and the real limits that decide whether you
Google Cloud made Cloud Run cross-region failover GA in July 2026, days after a 15-hour Netherlands outage. Here is a production multi-region setup, the gcloud commands, the real cost, and the data-tier caveat most
A single hour of downtime can cost a large enterprise over $300,000. eCorpIT designs multi-region resilience and disaster recovery: the right DR tier per workload, automated failover, and drills that prove your RTO
AWS Bedrock AgentCore, Google Gemini Enterprise and OpenAI Presence define three ways to run enterprise AI agents. We compare pricing, governance and lock-in as of July 2026 to help CTOs choose.
The price war we flagged in June arrived. On 30 July 2026 OpenAI cut GPT-5.6 Luna by 80% and Terra by 20%, and gave Sol a faster mode. Here are the new numbers, the competitive backdrop with Anthropic, and how the cut
AWS launched a free FinOps Agent in June 2026 on Bedrock. A 2026 decision guide comparing it with Finout, Amnic, Cloudability and native Cost Optimization Hub, with pricing, so you know when native is enough and when
On June 8, 2026 AWS added target coverage analysis to the Savings Plans Purchase Analyzer. Here is how to pick a coverage percentage, run the analysis, and size a commitment without paying for unused capacity.
Cloudflare Workers bills $5/month plus requests and CPU with no egress; Vercel Pro is $20 per seat plus metered compute and transfer. A 2026 cost breakdown with worked monthly examples.
CPU-based autoscaling scales GPU inference too late. A 2026 config guide to HPA, KEDA, external metrics and Karpenter, with YAML for queue-depth and scale-to-zero scaling of LLM inference on Kubernetes.
Frontier models cluster within points on every eval, and your workload reorders them. A senior engineer's scorecard for choosing an LLM in 2026: predictability, latency, token economics, and supervision load.
On July 21, 2026 AWS split Amazon SES into three plans, Essentials, Pro and Enterprise, while keeping a-la-carte pricing. The plans are not cheaper per email, so the real decision is tooling, not send price.
AWS reduced ECS Managed Instances management fees 60% for P-series and Trainium and 35% for G-series from July 1, 2026, the same day Capacity Block GPU rates rose. Here is how to capture the cut, with configuration
Claude Sonnet 5 launched on June 30, 2026 at introductory $2/$10 pricing through August 31. On September 1 it moves to $3/$15, and a new tokenizer already counts about 30% more tokens for the same text. Here is the cost
On July 6, 2026 Databricks Genie moved to pay-as-you-go DBU billing beyond a free per-user allowance. Here is how to set budgets, per-user caps and a billing query so a Genie rollout never surprises finance.
Databricks Genie meters text-to-SQL in DBUs from July 6, 2026; Snowflake Cortex Analyst bills per message at a flat AI Credit price. A 2026 comparison of which is cheaper to run and easier to forecast.
Gemini charges thinking tokens at the output rate, which is why a reasoning-heavy coding session on Gemini 3.1 Pro can burn cash fast. Since the free Code Assist and CLI tiers ended on June 18, 2026, that spend is now
Grafana Cloud added Agent Observability for LLM agents, built on OpenTelemetry. This guide covers instrumentation with OpenLIT, what the traces show, and the trace-volume cost math versus Datadog, Langfuse and Arize
Kimi K3's API costs $3/$15 per million tokens. Self-hosting the 2.8T open weights needs roughly 1.4 TB of VRAM. We size the GPUs, price the node against 2026 cloud rates, and work out the break-even.
Amazon Bedrock AgentCore moved every agent trace, prompt and log into a single per-agent CloudWatch log group and lifted default limits to 5,000 concurrent sessions. A config-level guide to the July 2026 observability
Three budget-tier models now sit far below the flagships: Gemini 3.5 Flash-Lite at $0.30/$2.50, GPT-5.6 Luna at $1/$6, and Claude Sonnet 5 at $2/$10 per million tokens. A worked cost comparison for high-volume
GitHub Copilot moved to usage-based AI Credits on June 1, 2026 and Cursor split its seats into usage pools. That reshapes the BYOK-versus-subscription math for every dev team. A worked cost model and a decision guide.
A cost breakdown of the Gemini Enterprise Agent Platform in 2026: the seat editions, the per-million token rates, and the July-to-September billing dates for Skills Registry, Agent Gateway, Memory Bank and Sessions that
Yotta's 85,000-GPU buildout and IndiaAI's Rs 65 per hour compute make domestic Blackwell cloud real in 2026. A decision framework weighing price, data residency, latency and availability against the hyperscalers.
AWS launched Database Savings Plans at re:Invent 2025: up to 35% off nine managed database services for a 1-year commitment. We run the real numbers against Reserved Instances so you know when each one actually wins.
AWS shipped two NVIDIA Blackwell inference families in 2026: G7 with up to 4.6x the inference of G6, and G7e that runs a 70B model on a single 96 GB GPU and is now live in Mumbai. Here is the spec, cost and decision
AWS Trainium3 is generally available in 2026 with 30-40% better price-performance than Trainium2. Here is how it compares with NVIDIA H200 and B200 on real cost per token and what Neuron migration costs.
India's public cloud spending is forecast to grow 28.1% to $17.5 billion in 2026, and Flexera says wasted spend rose to 29% for the first time in five years. This is a refreshed, senior-engineer FinOps playbook: nine
CloudWatch Coding Agent Insights, launched July 20, 2026, gives engineering leaders OpenTelemetry dashboards for AI coding tools: how to measure token spend, per-team cost and developer productivity across Claude Code
A specs-and-cost comparison of Google Ironwood (TPU7x) and NVIDIA B200 for AI inference in 2026: memory, bandwidth, scale, software lock-in, consumption models, and a decision matrix for when to pick each.
India's data centre numbers get quoted interchangeably, but 1.9 GW of capacity, $120 billion in commitments, and a $1.7 billion market are three different things. Here is how to read them before you decide where to host
AWS's June 2026 container caching removes the ECR image pull from SageMaker scale-out. In AWS's own test a GenAI endpoint's cold start fell from 525 to 258 seconds. The benchmarks and when to use it.
Over 40% of agentic AI projects will be cancelled by the end of 2027, Gartner says, on escalating cost, unclear value and weak risk controls. A build plan for staying in the surviving 60%.
On July 22, 2026 Cursor launched Router, a request-level classifier that sends each coding request to the best model. In Cursor's own tests it cut costs 30-60% versus Opus 4.8. Here are the numbers and the trade-offs.
AMD put EPYC Venice into production on TSMC 2nm in May 2026, then spent July announcing where it lands first: two new Azure VM series and Anthropic's 2 gigawatt Helios deployment. Here is how to read the numbers before
S3 Standard runs $0.023/GB and GCP Standard $0.020 in 2026, but egress and retrieval fees decide the real bill. Compare AWS, Azure and GCP storage tiers, and the tactics that actually cut spend.
Alibaba Cloud unveiled Agent Native Cloud at WAIC on 18 July 2026, adding AgentLoop and AgentTeams to AgentRun. Here is how the three big agent runtimes compare on billing, isolation, observability and lock-in.
AMD and NVIDIA both ship 72-accelerator AI racks in 2026, and almost none of their headline numbers are measured the same way. A line-by-line normalisation of the two vendor-published spec sheets, and the seven
On 16 June 2026 Google Cloud changed the default CUD scope from Project to Billing account. Commitments bought by one team now discount usage in projects that never paid for them. Here is how to check your scope
Amazon EKS 1.33 drops out of standard support on 29 July 2026 and billing jumps from $0.10 to $0.60 per cluster hour that same day. Here is the upgrade path to 1.36 and every breaking change between.
Amazon RDS for MySQL 8.0 reaches end of standard support on 31 July 2026, and Extended Support billing begins the next day at roughly $292/month for a single db.r5.xlarge. Azure and Google Cloud bill from 1 January
Gartner expects power to constrain 40% of AI data centers by 2027, and GPU lead times run 36 to 52 weeks. This 2026 playbook covers capacity forecasting, spot vs reserved, neo-clouds and procurement governance.
A 2026 cost-per-token comparison of the NVIDIA B200 and H100 for LLM inference, with MLPerf v6.0 throughput, real cloud hourly rates, the arithmetic that makes the pricier chip cheaper per token, and the cases where
Frontier LLM output prices fell about 94% since 2023, and a 100x spread now separates the cheapest models from the priciest. How to migrate off deprecated models and cut your AI bill.
GitHub Copilot now bills by GitHub AI Credits, not seats alone. A senior engineer's guide to credit pools, cost-center caps, per-user budgets and chargeback that holds at scale in 2026.
Meta reportedly plans Meta Compute to rent out its spare AI GPUs, and the news moved billions on 1 July 2026. What is confirmed, why the neoclouds fell hardest, and whether renting surplus GPU capacity is worth it.
AWS Interconnect - multicloud reached GA on 14 April 2026 with Google Cloud as first launch partner. A 10 Gbps Tier 1 link costs $9,000.90 a month with no per-GB charges. Attach a Cloud WAN Core Network with a Singapore
Azure Update 559756 paused the 128 KiB minimum billable object size for cool, cold and archive tiers. Billing did not change on 1 July 2026 for new or existing accounts. AWS still bills a 128 KB floor; Google does not.
Cursor split Teams usage into two pools and added a $120 Premium seat for renewals from 1 July 2026. Premium buys 5x the included usage at 3x the cost, which makes its included usage 40% cheaper per unit. Here is
Thinking Machines released Inkling on 15 July 2026 under Apache 2.0: 975B parameters, 41B active. We priced the GPUs against the Tinker API using AWS and vendor figures, and the break-even is brutal.
On July 1, 2026 Azure stopped new purchases and renewals of Reserved VM Instances for 18 VM series, including Dv3, Ev3, Av2 and Fsv2. Existing RIs still run to term. Two options remain, the renewal escape hatch is gone
Anthropic began showing rupee prices in India on July 13, 2026. Claude Pro lists at ₹2,000/month against $17 in the US. Most of that 24% gap is GST, not an India surcharge - and the plan-versus-API decision matters far
Egress is the cloud line item nobody quotes in the sales deck. Here is what four providers charge per GB as of June 2026, the hidden charges that dwarf the headline rate, and what the EU Data Act actually abolishes
In-place pod resize graduated to stable in Kubernetes v1.35, and pod-level resize reached beta in v1.36. What changed, what still forces a restart, and how to wire it to the Vertical Pod Autoscaler.
Three frontier models shipped within ten days of each other in mid-2026, at list prices that differ by 5x. The published price is not what you pay. Here is how to route workloads on evidence instead of vendor loyalty.
Microsoft's July 2026 price update reads as a spread of increases from 5% to 43%. Read the table as dollars and it collapses into one number: every enterprise suite rises $3, and the SKUs without Teams take the same hit
AI infrastructure spending is projected at $1.5 trillion for 2026, and Sequoia's arithmetic says the industry must earn $3 trillion to justify it. Frontier lab revenue sits in the low hundreds of billions. Here is what
AWS lifted EC2 Capacity Block rates roughly 20% on July 1, 2026, six months after a 15% rise on January 4. P6-B300 now bills at $14.04 per accelerator hour. Here is what changed, why analysts disagree on the cause
vLLM's April 2026 benchmark put FP8 KV-cache 14.9% above BF16 on one H100. Priced out, that is $4.24 down to $3.69 per million output tokens - and your AWS purchase model moves the number four times further.
Kubernetes is the default home for AI, yet GPUs often run at 5% utilization, wasting over $100,000 a year. How a managed platform fixes GPU scheduling and cost for Indian enterprises running AI.
Microsoft expanded its AWS-to-Azure migration tooling in 2026. But migration has hidden costs: moving 50 TB can cost $3,500 to $7,000 in egress alone. How to evaluate a switch on total cost, not vendor marketing.
About 60% of cloud migrations deliver below their projected ROI, and hidden costs can inflate a bill by 200-300%. Here is how to modernize legacy stacks in 2026 without blowing the budget, and how eCorpIT delivers it.
Bloomberg reported on July 1, 2026 that Meta is building a cloud business to rent out AI compute and host Llama models. A fourth serious cloud vendor changes the math on price, negotiation, and lock-in for enterprise
NVIDIA's Vera Rubin NVL72 is in full production and promises up to 10x lower cost per token than Blackwell, arriving in the cloud in H2 2026. Here is why your GPU bill may still climb first, and the FinOps moves that
Cloud waste hit a five-year high of 29% in 2026 as AI workloads made spend unpredictable. eCorpIT's cloud FinOps managed service brings visibility, rightsizing, and anomaly detection to AWS, Azure, and GCP so teams cut
AWS's FinOps Agent entered public preview at FinOps X 2026: agentic AI that root-causes cost spikes, answers plain-language cost questions, and attributes Bedrock AI spend by model and user.
In 2026, 98% of FinOps teams manage AI spend, Azure Copilot ships optimization agents that act under governance, and Graviton4 lifts price-performance up to 40%. Here is what cloud teams and CFOs should automate
The majority of enterprise AI agent pilots never reach production, and Gartner expects over 40% cancelled by 2027, mostly for cost and governance reasons, not model quality. Here is how eCorpIT builds agents that ship
OpenAI's GPT-5.6 (Sol, Terra, Luna) went public on July 9, 2026, with Luna at $1/$6 per million tokens. We compare the new tiers to Claude Sonnet 5 and Gemini, and the levers that cut inference bills 60 to 80 percent.
In the 2026 State of FinOps, 98% of teams now manage AI spend and GPU cost has become the top concern. Here is what changed and how enterprises can regain control of the AI cloud bill.
LLM API prices run from $0.14 to $30 per million tokens in June 2026, and agentic workloads multiply the bill. Seven free, open-source tools that make engineering teams' LLM spend measurable and attributable.
With 98% of FinOps teams now managing AI spend, LLM cost control is a 2026 priority. Here are 6 free, mostly open-source tools, from LiteLLM and Langfuse to tokencost and OpenCost, to measure LLM spend.
Strategic optimisation can cut LLM costs by 60 to 80 percent. Seven AI cost tools, from observability with Langfuse to prompt compression with LLMLingua, that engineering teams use to cut token spend in 2026.
Companies waste about 27% of cloud spend, and AI is now the fastest-growing line. Six FinOps ways Indian teams cut AWS, Azure and GCP bills in 2026, from commitment discounts to cutting the GPU and token bill.
