Nvidia board reportedly prepares to oust Jensen Huang

Welcome to The Signal from Superpower Daily with Maya and Theo

Today

we start with a massive potential shakeup in the AI hardware world as NVIDIA's board reportedly prepares to oust CEO Jensen Huang

Right

And the stakes here really go far beyond a standard corporate power struggle

NVIDIA is essentially the physical bottleneck for the entire global tech economy

Yeah their GPUs are the absolute backbone of the AI ecosystem

I mean

they power the training of these massive models Exactly They

power the training and they also power inference And inference

is the process of running those models at scale every single time a user types a prompt Well the

Wall

Street Journal broke this story They report

that NVIDIA's directors are preparing to determine his future at the company And that

board meeting could happen as soon as Wednesday November 11

We really

have to look at the history to understand the gravity of this Jensen Huang

co founded NVIDIA He has led the company for 33 years Wow

33

years Yeah

He

built it from a niche graphics card manufacturer for video games into a multi trillion dollar infrastructure monopoly

Which is incredible

It is

So removing a founder with that much embedded knowledge is incredibly risky

His management style is also notoriously unique for a company of this scale

He reportedly has about 60 people

reporting directly to him That

is an intensely flat corporate structure I

mean most CEOs of Fortune 50 companies have maybe 8 to 12 direct reports Right

8

to 12 standard Exactly So

60

direct reports means Jensen centralizes an immense amount of operational control in himself He is

involved in the minute details of product roadmaps and supply chain negotiations And TechCrunch

reports that he delegates very little Yes That

creates

a massive bus factor If you

remove the single node connecting 60 different departments the entire

decision making apparatus basically freezes The

structural risk is enormous There

is no obvious successor here NVIDIA

has no clear succession plan in place None

at all None

at all Apparently

not If

the board does remove him there

is no designated executive ready to step into that central role The

company would be paralyzed And

at the exact moment their competitors are trying to chip away at their market share Precisely

Now

let's talk about the market reaction Right

TechCrunch

noted an 8 decline in after hours trading directly following the news They also

noted an 11 stock decline over five days The market

absolutely hates uncertainty They really

do Wait though

I am

looking at this 11 drop That five

day window includes trading before this report even broke That

is true We

cannot pin all of that decline on the ouster news The

stock had already started sliding earlier in the week You

make a good point The

broader 11 decline reflects a wider market anxiety Tech

stocks have been volatile lately Yeah

they have But

the 8 after hours drop is the specific direct reaction to the journal's reporting

We have to separate the immediate shock of the board rumors from the longer macroeconomic trend

That makes sense

We also need to be clear about the actual status of his employment

TechCrunch noted that the company has not announced a dismissal

Right

It

is a reported preparation for a discussion It

is immediately responded to CNBC for

comment The supply chain implications are really what you

should care about if you are building in this space Entire industries depend on NVIDIA's execution Absolutely Data

center construction schedules literally depend on

their

chip deliveries Foundational model release dates are tied to access

to NVIDIA's hardware So a disruption at the very

top ripples out to every

founder and operator waiting for compute Exactly The immediate question is whether the directors actually meet on

Wednesday

We also need to see what official communications emerge from

NVIDIA regarding their leadership stability Until the board makes a formal statement the market is

operating on pure speculation And the silence

from the company is just adding fuel

to that fire Definitely Well that wraps up our look at NVIDIA for

now

We

will watch for their official response Next up we

are moving from the physical hardware bottleneck

to the

software layer where this compute is actually applied Google DeepMind has released Alpha Genome 2 Right This is

a new AI model that shifts

genetic

predictions from broad tissues to specific cells This fundamentally well it changes how we use AI

in biology

DeepMind released this model on November 11th Okay To

understand why this matters you have to

look at how DNA actually

works Your heart your brain and your immune system all share

the exact same DNA sequence Right But a genetic variant might cause a

fatal disease in one cell type and do absolutely nothing in another I see So the original

Alpha Genome model

analyzed a

million letter DNA sequence and predicted molecular properties across thousands of experiments and tissues all at once Yes it did But DeepMind says that broad view

spread

the model's

capacity too thin It blurs subtle effects that only appear in specific

cells Broad tissue averages really obscure the microscopic details If

you blend a thousand different cell types together

and analyze the average you miss the one specific immune cell that is

actually malfunctioning Alpha Genome 2 fixes this by conditioning its predictions

on a

chosen biological context It pairs the DNA sequence with a specific tissue cell type

or cell state and it feeds that specific condition

into every

single stage of the

model The mechanism is highly targeted So instead of taking a national poll

to predict behavior this model

asks how a very specific neighborhood behaves on a very

specific day That is a great analogy It is exactly like that DeepMind trained

this model on an immense

amount of data to achieve

that zoom in effect The training combined bulk measurements with single cell data from human

and mouse samples Right They are talking about more than 107 million individual

cell

profiles across 68 tissues That is massive It

is And the results on an

immune cell benchmark

called 10k10k

show a huge leap in capability This benchmark covers 28 primary immune cell types DeepMind reports

that Alpha Genome 2 doubled its causal variant

detection It identified 30 of causal variants compared to 16 for

the prior model And identifying causal variants is basically the holy grail for

drug discovery A causal variant is the

actual genetic mutation driving a disease

not just a random mutation sitting nearby Exactly The directional predictions

also improved dramatically The model correctly identified

whether

a variant increases or decreases gene

activity 43 of the time And the previous model only managed 23 Right Understanding the directional

impact tells researchers whether they need to

invent

a drug to boost a gene's activity or suppress it

DeepMind tested this on a known inflammatory bowel disease mechanism The model computationally reproduced

a finding that a distant DNA switch increases activity of

the ETS2 gene in specific gut immune cells Yes It singled out the causal variant and its selective effect But we do

have

to state a vital limitation here The model reproduced an

existing finding It did not discover a new disease mechanism

That is a very important caveat

These impressive metrics are company reported research benchmark

results They are not verified clinical

detection rates Right This model is not validated for

clinical or diagnostic use Still the model

can

handle experiments it has never seen before Researchers can give

the

model gene activity data from just a few dozen cells

A cell encoder compresses that into a mathematical representation of the cell state Then

the model can predict how DNA behaves in that unseen cell state without

needing a completely new training run DeepMind actually demonstrated this

using unseen data from a COVID 19 mRNA vaccine study Oh really

Yeah The model successfully distinguished activated immune T cells from resting ones It

predicted that

a variant linked to severe COVID 19 disrupts gene activity much more

strongly in the activated state And if a researcher only looked at unactivated cells in a petri dish they

would miss that connection entirely Exactly Now the Alpha Genome API is available

for non commercial research Commercial access

is

currently in private preview The transition from research benchmark to

commercial pharmaceutical application is definitely the next

critical hurdle We should watch how pharmaceutical companies utilize this in those private previews We

absolutely should Meanwhile OpenAI's third quarter revenue has reportedly jumped about 50

to roughly 10

billion

While its adjusted operating loss has also widened These figures reveal a

massive structural shift in how AI companies make

money OpenAI hit roughly 10 billion in Q3 revenue That is a sharp

acceleration

from the 6 7 billion they reportedly recorded

in Q2 The top line revenue growth is staggering for a company this young

But the costs are compounding just as fast They

really are The adjusted operating

loss increased from 10 1 billion in Q2

to 10 9

billion in Q3 The adjusted operating margin did improve though It moved from negative 150

to negative 109 Well a better

margin

simply means sales are growing faster than the

losses are expanding The ratio is improving I am highly skeptical of celebrating an improving

negative margin at this

scale though I mean when does the sheer gravity of a 10 billion quarterly

cash loss become physically unsustainable It is a fair question The actual dollar loss still exceeds their

total quarterly sales The dollar

amount being burned is historic And the true financial picture

is even more complicated when you look

at the accounting How so The 10 9 billion loss explicitly leaves out 2 6

billion in stock based compensation Oh wow Yeah And if you include

the stock

payouts

they'll hole is much deeper OpenAI President

Greg Brockman noted where this

revenue surge is actually coming from Enterprise revenue doubled in the quarter It

now accounts for the majority of

the

company's sales The financial architecture of the leading AI company

is shifting heavily toward business to business sales OpenAI is essentially subsidizing massive losses

to become the invisible utility grid for enterprise businesses That makes sense They are lowering

prices on their

new models They are pushing hard for developer

API usage This indicates a fierce market fight

against competitors like Anthropx Cloud They want companies to build internal

tools on their infrastructure Once a Fortune 500 company integrates

the

OpenAI API into their daily routing and data processing they are locked in Exactly Now the

consumer base remains massive

They

reportedly have 1 2 billion weekly

active users and 100 million paying subscribers But a large consumer audience has

not brought operating costs below revenue Consumer subscriptions alone cannot cover the

compute bill No they cannot Bloomberg also reported

a 50 billion

annual revenue run rate Right but that figure just projects

a

recent sales pace forward It has not completed yearly revenue

Correct And none of these specific

figures

have been officially confirmed by OpenAI The compute costs for inference are

the root cause of this cash burn Every time an employee at an

enterprise client queries the API it costs OpenAI physical server time

An enterprise scale means massive continuous

query volume The thing to watch next is

whether enterprise sales growth can physically outpace the compounding costs of

those AI operations The reported margin is improving on

paper but the actual quarterly loss is

still moving the wrong direction Right The existential math for OpenAI

is

whether they can drive the cost of compute down faster than their enterprise clients scale up their usage In other news Google has launched

its Gemini Decisions

API This allows developers to score up

to 1 000 bounded questions against a single shared

input This completely changes the cost structure

for automated reasoning The new Gemini Decision 1 model

evaluates candidate options in a single processing pass It does not generate conversational text

token

by token We really need to break down why

that matters Standard generative AI predicts the next word

in a sequence It is like writing a customized letter

from scratch every single time you ask a question That requires massive compute power It does

The

Decisions API functions more like

a bouncer

at a club checking a VIP list It accepts yes or no checks multiple choice

questions or rubric scoring That is

a great way

to put it

The model simply returns probabilities and a confident

score for each option It does not waste compute writing out a

pros answer And the billing structure makes this highly disruptive

Inputs cost 0 10 per million tokens Crucially

the shared multimodal input is only billed

once

You can upload a one hour customer service

call recording and an entire PDF policy manual You pay for that input once You

can then ask the model 500

different

yes or no questions about that call Did the agent mention a refund Was the

customer angry And each additional question

only adds its own

specific text instructions to the input bill There are no output token charges Exactly Does

this signify an industry pivot I

mean

are we moving from AI as

a conversational chatbot to AI as invisible structural routing

software I think we are Bounded decisions are the foundation of

enterprise automation The confidence

scores range from 0 to 1 This allows

developers to set thresholds They can automatically send

low

confidence ambiguous decisions to another system or to a human reviewer

And the high confidence decisions

get routed instantly Google claims

this makes it incredibly cheap They report

it

is nearly five times cheaper than Gemini

3 5 Flashlight for these specific tasks across their benchmark suite Google's own Gemini notebook already uses this model to route

everyday queries They say it cuts latency by as much as 40

internally It quickly decides which deeper reasoning model should actually handle

complex tasks rather than wasting a large model on a simple sorting question There are major caveats to their advertised speed though Google

reports an end to end latency of 150 to

250 milliseconds They say speed stays nearly flat from 1

to

100 questions But that benchmark is based on contexts of up

to 16 000 tokens And the API actually accepts a maximum input

of 256 000 tokens Google's not guaranteeing that lightning fast speed

at maximum capacity The speed claim operates within a

very defined

envelope What listeners

should watch next is how real world latency holds

up As developers push this API closer to its 256 000 token limit in live production environments see if the speed remains viable Routing software has

to be instantaneous to be useful

If pushing the context window slows the routing down

to 3 or 4 seconds it loses its primary advantage over traditional code That brings us to our Quick Reads section Starting

our Quick Reads with Aviation United Airlines is

moving to block unauthorized third party

personal AI agents from accessing its booking systems The airline announced this policy update on November

11th United is blocking AI bots from booking flights

And this applies even when these unauthorized agents

route their requests through travel agencies or aggregators It does United cites security and regulatory risks

They warn

of stolen mileage plus miles unauthorized account changes and exposed personal

information They also

raise federal aviation compliance

issues Right Automated check in by a

bot could bypass a human acknowledgment that checked bags contain no explosives That is a huge

safety issue It is Aviation analyst

Henry Hardevelt

notes this is also about business control United wants to maintain

ownership of the booking process And they also want

to avoid the massive server costs generated by constant bot traffic scraping their site Exactly The airline has

not

explained exactly how it will identify these unauthorized agents They also have not

stated when enforcement will begin Watch for their upcoming

Digital Trust Center update for details on how they plan to police this Next up in Workplace

Tech XAI has brought Grokbot

into Google Meet Zoom and Microsoft Teams to

take notes and answer questions in the meeting chat This is part of

the Grokbot 1 0 release The agent can

join calls transcribe the conversation

and actively answer

questions in the text chat pane And its transcribing

label is visible to all participants It is But

the UI

controls present a significant security caveat Disabling interactivity stops

the bot from answering questions in the chat It does not

stop

the bot from listening and taking notes Muting its answers does

not end its access to the meeting audio You must explicitly

hit the Quit button

to remove its access Hitting Cancel only removes a meeting that is still scheduled Watch how enterprise administrators utilize the new

dashboard switch They have the power to block bot

attendance entirely at the administrative level Finally in Consumer AI ByteDance is reportedly

preparing

an AI assistant

modeled on Meta's use for release within weeks ByteDance wants to release a multi step

agent that works autonomously across connected apps and browsers This aggressive product timeline contrasts sharply with

recent internal admissions Co founder Zong Yiming reportedly told his

AI team that matching OpenAI's advanced research models will take until 2029 or 2030 Zong pointed to a severe lack of pre training data

ByteDance's models use data sets that are orders of magnitude smaller than those of frontier competitors There is no confirmed release date or hardware pairing Watch

to see if a product built on smaller pre

training data sets can actually execute complex multi step actions reliably in the real world It will

be interesting to see We are moving to

three takeaways from today First a

massive vulnerability exists at the top of the

AI

hardware chain A single leadership crisis at NVIDIA immediately shakes the market

proving how dangerously centralized the physical AI boom really is

Second the economics of AI software are fundamentally shifting OpenAI

and Google are aggressively reshaping their pricing and

API structures to capture enterprise routing This proves the real money is in invisible business

infrastructure not just consumer chatbots Third we are watching a

live clash between AI

convenience

and system security From United Airlines blocking booking bots to XAI's complicated

meeting permissions companies are struggling to govern autonomous agents interacting with

real world systems Tomorrow we will be watching for any official statements

from NVIDIA's board regarding Jensen Huang You can find more at SuperPowerDaily com Thank you

for listening We'll see you tomorrow

Thank you for listening.

We'll see you tomorrow.