AI Morning Briefing 2026/08/27: Nvidia's Q2 Revenue Hits $96.2B, Anthropic Plans $45B AI Compute Lease, MiniMa…｜Article

Today's AI news marks the shift toward scale economies: Nvidia's Q2 revenue doubles to $96.2B with $89B from data centers; Anthropic reportedly plans a $45B, six-year AI compute lease with Nscale; Chinese AI firm MiniMax's first-half revenue jumps 283%. The next AI challenge is turning massive…

Here are three AI news items today.

On the surface, they talk about:

Chips.

Data centers.

Chinese AI company earnings.

But together,

they actually answer the same question:

Is the huge investment in AI finally paying off?

In recent years,

everyone's been competing on:

model parameters,

benchmarks,

context windows,

and inference power.

Now the question turns more practical:

Who is really buying all these GPUs?

How much compute do model companies actually need?

Are cheaper models truly generating paying customers?

These three stories together

connect all three crucial layers

of the AI industry.

First: Nvidia's Q2 revenue $96.22 billion, data center revenue $89 billion in one quarter

Nvidia just released

its fiscal 2027 Q2 report.

Quarterly revenue:

$96.22 billion.

This is

a 106% increase over last year,

meaning it more than doubled within one year.

The most impressive figure is:

Data Centers.

Quarterly revenue:

$89 billion.

Year-over-year growth:

117%.

This indicates AI data center demand

has not disappeared despite years of heavy investment.

Instead, it's

still very robust.

More importantly: Nvidia is no longer reliant on just a few large AI customers

Jensen Huang said in the earnings call,

one year ago the market was mainly driven by

a handful of large Frontier Labs.

Now it has shifted to:

many leading AI labs,

a large number of startups,

open models,

and Physical AI,

all expanding simultaneously.

This matters greatly for Nvidia,

because one of the biggest risks for chipmakers is

over-concentration of demand.

If the entire AI boom depended on only one or two companies

constantly buying GPUs,

the risk would be significant.

Now Nvidia wants to prove

the demand base is broadening.

The next-generation Vera Rubin chip is already in full production

Nvidia stated,

the next generation,

Vera Rubin,

is already in

full production.

This means Nvidia isn't waiting until

Blackwell demand ends to decide their next move,

but the next product line

is already in place.

For the AI infrastructure industry,

the key question is whether customers are willing to

keep upgrading

generation after generation.

Nvidia projects no slowdown in the next quarter

The company estimates Q3 revenue around

$108 billion.

The market previously expected

about $104.2 billion.

This means Nvidia not only beat expectations this quarter,

its outlook for next quarter

also exceeds market estimates.

This is why Nvidia's stock price

rose after hours following the earnings release.

But the most important point today isn't "Nvidia making a lot of money again"

The real signal is:

AI compute has truly become a giant industry.

Jensen Huang even said outright:

Compute is revenue.

Simply put,

people used to view

computation as a cost.

Now, more and more AI companies see compute as directly determining

how many users they can serve,

how many agents they can run,

how many tokens they can sell,

and ultimately

how much money they can make.

Second: Anthropic reportedly plans $45 billion six-year compute lease

If the first story shows

someone is really selling huge amounts of GPUs,

the second answers

who is willing to spend that much?

One answer is:

Anthropic.

According to Reuters citing Bloomberg,

Anthropic is reportedly preparing to

lease massive compute resources from

AI infrastructure company Nscale,

with a total value of about

$45 billion

over approximately

six years.

This averages about $7.5 billion per year

This is just a

compute lease agreement,

not:

an acquisition,

not

a merger,

not

employee expenses.

It's spending to

keep Claude running.

This alone explains

what truly expensive frontier AI is now.

It's not

making chatbot interfaces,

but maintaining

the entire massive compute infrastructure behind it.

Reportedly sourced from large West Virginia data centers

Reuters says the compute will come from

Nscale’s data center campus in West Virginia, USA,

using Nvidia’s

Vera Rubin systems.

Important to note:

Anthropic has not officially confirmed this deal.

Reuters says Anthropic declined to comment.

So currently the correct phrasing is

"reported to be arranging/planning"

rather than

"Anthropic has officially signed a $45 billion contract."

Why does Anthropic need such massive compute?

Because Claude is no longer

a simple chatbot.

Anthropic’s revenue increasingly comes from

enterprise,

API usage,

Claude Code,

and agent work.

These services share one thing in common:

Greater usage means greater inference compute requirements.

If an engineer uses Claude

ten times a day,

versus letting Claude Code

continuously program for an entire company,

the compute demand is on a completely different scale.

AI agents could make compute demands even more intense

A regular chatbot:

You ask once,

it answers once.

An agent might complete a task by

thinking,

searching,

using tools,

reassessing,

and acting multiple times on its own.

One user clicking

"complete this task"

might trigger dozens or even more model computations behind the scenes.

If agents really take off,

the biggest beneficiaries won’t just be software companies,

but also

compute sellers.

This also explains why AI companies are aggressively locking in long-term compute contracts

If demand skyrockets,

the worst case is

models have customers but

lack GPUs,

data centers,

or electricity.

So companies like OpenAI,

Anthropic,

Google,

and Meta

are competing not just for model talent,

but also

future years’ physical compute supply.

But this $45 billion figure raises another question

Can revenue keep up?

Compute is not free.

Data centers are not free.

Electricity is not free.

If AI companies spend tens or even hundreds of billions annually

on compute leases,

there must be

enough customers

to keep paying.

This is why

AI company revenue figures

are becoming ever more important.

The third story today

precisely addresses this.

Third: MiniMax’s first-half revenue up 283.1%, surpassing 2025 full-year revenue

Chinese AI company

MiniMax

released its H1 2026 financials.

First-half revenue:

$116.6 million.

283.1% year over year.

Interestingly,

MiniMax’s total revenue for all of 2025 was only around

$79 million.

This year, in just half the time,

it already surpassed the entire previous year.

The real explosion is not consumer chatbots but enterprise and API revenue

MiniMax’s

Open Platform

and other enterprise AI revenue rose from

$9.2 million last year

$73.9 million.

A growth of

703.1%.

It now accounts for

63.4%

of MiniMax’s total revenue,

compared to 30.3% a year ago.

This is a major structural shift.

This means MiniMax is moving from an "AI app company" to an "AI foundational service company"

MiniMax was previously known mostly for

Talkie,

Hailuo AI,

video generation,

and general AI apps.

But the fastest growth now comes from

enterprise,

developers,

APIs,

and token plans.

This means other companies are embedding MiniMax’s models

into their

products and workflows.

This kind of revenue generally shows

whether a model is truly in production,

much more than a casual user subscribing to an AI app.

AI-native products are also still growing

MiniMax’s AI-native product revenue in H1 was

$42.6 million,

100.9% year over year.

So MiniMax is not abandoning consumer products

to focus only on APIs,

but both revenue streams are growing,

with enterprise and APIs growing faster.

MiniMax also shared an intriguing usage statistic

The company said by July 2026,

MiniMax token usage

was

20 times higher than one month ago.

So in just over half a year,

usage increased 20-fold.

This also explains why all AI companies talk so much about

inference efficiency.

Because if usage rises 20 times, costs can bankrupt a company without efficiency gains

Assuming

each million tokens

has a fixed cost,

a 20-fold increase in usage

means inference compute costs

would soar accordingly.

MiniMax continually stresses

Performance-Cost Frontier.

They aim not just for

stronger models

but for equivalent capabilities

at lower costs.

Because once AI is used at scale,

cost becomes the key factor

for whether a business turns a profit.

MiniMax’s gross profit and margin have improved significantly

In H1,

gross profit increased from

$3.7 million last year

$20.81 million,

a growth of

464.8%.

Gross margin rose from

12.1%

17.9%.

This indicates

that as revenue grows,

infrastructure efficiency is also improving.

But let’s not rush to say MiniMax is "profitable" yet

It’s not.

MiniMax’s H1 adjusted net loss was

$293 million.

R&D expenses totaled about

$296.9 million.

Meaning, while revenue grew rapidly,

spending on R&D and model competition

remains substantial.

This is the key contradiction underlying the three stories today.

AI now generates real revenue, but investment needs are skyrocketing

Nvidia shows customers are

indeed buying compute.

Anthropic reportedly plans to spend

$45 billion

to secure compute in advance.

MiniMax proves that low-cost models and APIs

are starting to generate rapid

revenue.

So it’s no longer accurate to say

AI is just

hype

without customers,

because customers are paying.

But neither does that guarantee

all AI investments

will pay off.

The real next battle is "Unit Economics"

Simply put:

Does doing an AI task make a profit?

For example, an AI agent

may charge a customer

$10,

but behind the scenes

GPU,

electricity,

model inference,

searching,

and tool usage

could cost

$12.

More users could mean

larger losses for the company.

So don’t just look at revenue growth

Also consider:

per-inference cost,

gross margin,

customer retention,

R&D spending,

capital expenditures,

compute commitments,

and whether

these investments can ultimately

generate ongoing

paid usage.

Nvidia’s earnings are essentially the receipt at the top of the chain

When Nvidia reports quarterly revenue of

$96.2 billion,

it means

someone paid that money.

Those payments could come from

Microsoft,

Amazon,

Google,

Meta,

OpenAI,

Anthropic,

AI startups,

governments,

and enterprise data centers.

The real question is

how these customers

will get their money back.

Anthropic’s $45 billion reported deal represents the next layer

Nvidia:

sells compute.

Nscale:

builds data centers.

Anthropic:

leases compute.

Then

Claude’s customers

pay Anthropic.

The whole chain requires

someone at the end willing to keep

paying continuously.

Otherwise,

all that massive investment

runs into trouble.

MiniMax proves the bottom layer: "Someone is actually paying"

Their most important figure today is not

283.1%,

but the shift in enterprise and API revenue share from

30.3%

63.4%.

This means AI capabilities are becoming

services enterprises use daily.

That is genuine

commercialization.

This also explains why "the cheapest model" suddenly matters a lot

When models first debuted,

usage was limited.

Spending a few more cents per query

wasn’t a big issue.

But if a company runs

hundreds of millions of model tasks daily,

each costing even 0.001 dollars more,

the total difference becomes

enormous.

When AI truly scales,

cost may be more important than

benchmark improvements of a fraction of a percent.

These three companies represent different layers of the same chain

Nvidia

sells:

compute.

Anthropic

buys massive amounts of

compute,

then converts it into:

Claude.

MiniMax

works to sell models

more cheaply to

enterprises and developers.

All three are bound by one factor:

usage volume.

Without real usage, AI infrastructure is just expensive hardware

A data center

doesn't make money simply because it's built.

A GPU

doesn't become revenue just because it's purchased.

A model

doesn’t generate paying customers just by topping benchmarks.

The thing that truly connects

chip,

power,

model,

and API

is one thing only:

users who actually keep using it.

So the real question today is not "Has the AI bubble burst?"

That’s a too binary question.

A better question is:

Which layers already have revenue?

Nvidia:

Clearly yes.

Which layers are still in early-stage betting?

Large-scale compute contract deals:

Clearly yes.

Which layers show signs of commercialization?

MiniMax:

Emerging.

Which layers remain most uncertain?

Many

model companies,

AI agents,

enterprise AI business models

with sufficient high gross margins

still need to be tested.

The AI industry is shifting from "Who dares to spend money" to "Who can make a profit"

In 2024 and 2025,

the market mainly focused on

how much was invested,

how many GPUs were bought,

and how big models were.

Now in 2026,

One takeaway this morning:

The AI boom so far

can no longer be explained by

"Everyone is just burning cash."

Because Nvidia’s

$96.2 billion quarterly revenue

is real,

and MiniMax’s

283.1% revenue growth

is also real.

Demand and revenue in AI

are indeed starting to grow.

But on the other hand,

Anthropic’s reported compute agreement

$45 billion

reminds us:

AI commercialization isn’t just about having revenue.

The real challenge is

whether revenue growth can keep pace

with skyrocketing costs in

chips,